population and con ict - the review of economic studies€¦ · ict might work through channels...

Review of Economic Studies (2019) 01, 1–?? 0034-6527/19/00000001$02.00

c© 2019 The Review of Economic Studies Limited

Population and Conflict

DARON ACEMOGLU

MIT, 50 Memorial Drive, Cambridge, MA 02142.E-mail: [email protected]

LEOPOLDO FERGUSSON

Universidad de los Andes, Cra 1 N 18A - 12, Bogota.E-mail: [email protected]

SIMON JOHNSON

MIT, 50 Memorial Drive, Cambridge, MA 02142.E-mail: [email protected]

First version received April 2017; final version accepted July 2019 (Eds.)

Medical innovations during the 1940s quickly resulted in significant healthimprovements around the world. Countries with initially higher mortality from infectiousdiseases experienced larger increases in life expectancy, population and subsequent socialconflict. This cross-country result is robust across alternative measures of conflict, and isnot driven by differential trends between countries with varying baseline characteristics. Asimilar effect is also present within Mexico. Initial suitability conditions for malaria variedacross municipalities, and anti-malaria campaigns had differential effects on populationgrowth and social conflict. Both across countries and within Mexico, increased conflict overscarce resources predominates and this effect is more pronounced during times of economichardship (specifically, in countries with a poor growth record and in drought-stricken areasin Mexico). At least during this time period, a larger increase in population made socialconflict more likely.

1. INTRODUCTION

The world’s population is forecast to rise from its current level of around 7.4 billion tonearly 11.2 billion by 2100. This growth will be unequally distributed around the world.“More developed regions,” as classified by the UN, are expected to remain at roughlythe same population level, 1.25-1.30 billion, over the next century. “Least developedcountries,” on the other hand, are projected to increase their population from 957 millionin 2015 to 3.2 billion in 2100. The UN forecasts that the population of Africa will risefrom its current level of near 1.2 billion to almost 4.5 billion within the century.1 Whatis the likely impact of a population increase on this scale?

These population changes are not entirely unprecedented. During the internationalepidemiological transition starting in the 1940s, largely due to the introduction of newdrugs, chemicals and public health measures, many relatively less prosperous countriesexperienced major improvements in health and longevity (Acemoglu and Johnson, 2007).

1. The data in this paragraph are from the UN’s 2017 long-term population projections,downloaded on December 15, 2018 from https://population.un.org/wpp (United Nations, 2015). Latestannual data are for 2015 and the projections run through 2100. According to the notes provided by the UNwith the downloaded data: “More developed regions comprise Europe, Northern America, Australia/NewZealand and Japan,” and “The group of least developed countries, as defined by the United NationsGeneral Assembly included 48 countries in January 2014: 34 in Africa, 9 in Asia, 4 in Oceania and onein Latin America and the Caribbean.”

1

2 REVIEW OF ECONOMIC STUDIES

These innovations led to large increases in population. For example, between 1940 and1980, the population of the poorest third of countries in our sample increased from around1.13 billion to 2.55 billion.2

In this paper, we exploit this historical episode to shed light on whether changes inpopulation have a major impact on civil wars and conflict. We focus on the populationincrease between 1940 and 1980, during which global health technology and conditionsimproved dramatically.

In our cross-country analysis, we exploit variation in population growth coming fromthe initial distribution of mortality from infectious diseases.

Most of the medical and public health breakthroughs in this period originated ina few industrialized countries and can reasonably be seen as exogenous to developmentprospects and likelihood of internal conflict in the rest of the world. Our identificationstrategy does not depend on when a particular country adopted better public healthmeasures or how effectively these measures were implemented. We also document thatthere were no differential trends in life expectancy, population, conflict or economicoutcomes between initial high and low mortality countries before the onset of theinternational epidemiological transition.

Countries with higher exogenous increases in population experienced more socialconflict after 1940. Across different specifications and alternative definitions of conflict,instrumented changes in population have a significant positive effect on various measuresof civil war and violent conflict.

The effect of population increase on social conflict is sizable. A rise in log populationof about 0.68 from 1940 to 1980, corresponding to the average change in population inour sample of countries, caused roughly 4.2 additional years of full-blown civil war in the1980s relative to the 1940s.

An important concern is that the effects of changes in mortality from various diseaseson social conflict might work through channels other than population. For example, thesemortality changes also impact the overall health of the population and the age structureand may lead to greater education for new cohorts, all of which potentially affect conflict.Although we explore the various channels and show that they do not appear to havea first-order impact on our measures of conflict and cannot account for our results,this concern implies that our instrumental-variables results ought to be interpreted withcaution.

Our preferred interpretation for the linkages between disease mortality, populationand conflict is via a Malthusian channel: growth in population unaccompanied byimproved productivity or economic opportunities intensifies conflict over scarce resources.We provide a number of results consistent with this channel, in particular showing thatpopulation surges increase conflict related to natural resources but have no effect onconflicts unrelated to natural resources, and that they raise conflict in slowly-growing—but not in rapidly-growing—countries.

The second part of the paper examines the experience in Mexico. In 1940, Mexicohad an income per capita about 25% of the US level, slightly above the more prosperousparts of Asia and about three times the income level of India and many parts of sub-

2. Because our study requires detailed death-by-disease data, our base sample contains some—butnot all—countries that were low income in 1940. Our sample has good coverage for most regions in theworld, with the exception of Africa. Given the incidence of violent civil conflict in Africa, this is animportant limitation, but one that we cannot overcome with the available data, since most sub-SaharanAfrican countries lack reliable data on causes of death disaggregated by disease dating back to the 1940s,and these data are essential for our empirical strategy.

ACEMOGLU ET AL. POPULATION AND CONFLICT 3

Saharan Africa. Most important for our purposes, Mexican municipalities experienceddifferential changes in mortality from malaria depending on their initial exposure (asin our cross-country design). This differential incidence of malaria within Mexico wasdue almost entirely to differences in climate and geography. Starting in the 1940s, aspart of the broader international epidemiological transition, Mexico embarked upon anambitious anti-malarial campaign. These efforts quickly brought down morbidity andmortality rates from malaria.

We find that increases in population across Mexican municipalities driven by thesuccessful anti-malaria campaign are associated with more conflict, in particular, withmore violent protests during the 1960s. Consistent with our Malthusian interpretation,the significant increases in violent protests are driven by those related to naturalresources. Moreover, these protests were more pronounced when there was also economichardship in the form of drought during the growing season.

Our paper is related to several strands of research. There is an extensive literatureon malaria and the effectiveness of anti-malaria campaigns from a health perspective.Prominent papers include Barlow (1967), Case and Paxson (2009), Hong (2007), Ashrafet al (2010), Lucas (2010), Fluckiger and Ludwig (2017) and Barofsky et al (2015). Thisliterature does not examine the potential impact of population and health on socialconflict.

Following contributions such as those of Collier and Hoeffler (1998, 2004) andFearon and Laitin (2003), scholars have emphasized poverty, inequality, weak institutions,political grievances and ethnic divisions as explanations for the outbreak and persistenceof civil war. Most of this literature does not fully address the possibility that reversecausality or omitted variables bias may be driving the observed correlations. Blattmanand Miguel (2010) conclude that “further cross-country regressions will only be usefulif they distinguish between competing explanations using more credible econometricmethods for establishing causality” (p. 8). Miguel et al (2004) use annual rainfall as aninstrument for income growth in sub-Saharan Africa, while Besley and Persson (2008)rely on plausibly exogenous international commodity price movements.

Although population has not been a prime focus in the economics of conflictliterature (see the survey by Garfinkel and Skaperdas, 2007), there has been a livelydebate on the effects of population pressure on violent conflict in other disciplines,including political science. For example, Homer-Dixon (1991, 1999) studies the connectionbetween population growth, pressure on environmental resources and conflict, and findsthat poor countries are in general more vulnerable to environmentally-induced conflicts.This view motivated Robert Kaplan’s famous 1994 essay, “The Coming of Anarchy”,which predicted mounting conflict around the world. However, the literature is far froma consensus, and several authors, including Richards (1996), have pushed back againstthis view. Consistent with the notion that population pressures contribute to conflict, Tirand Diehl (1998) estimate that population growth pressures have a significant impact onsome dimensions of military conflict, especially in poor countries. Hauge and Ellingsen(1998) find that factors like deforestation, land degradation and water scarcity, aloneand in combination with high population density, increase the risk of domestic (low-level) armed conflict. Urdal (2005) does not find a strong correlation between populationgrowth and conflict risk, but confirms that conflict is more likely when high populationgrowth combines with land scarcity (see also Diehl and Gleditsch, 2001).

To the best of our knowledge only Bruckner (2010) attempts to establish thecausal impact of population size on conflict. He exploits randomly occurring droughtsas an instrument for population to address endogeneity and finds that a 5% increase in


population size raises the risk of civil conflict by around six percentage points. However,Bruckner’s study focuses on 37 sub-Saharan countries between 1981 and 2004, where theeffects of population pressures or droughts may be different than in other settings.

In Section 2 we present a simple motivating model where a Malthusian mechanismlinks population to conflict. Section 3 describes the data. Section 4 presents our cross-country analysis, including extensive robustness checks. Section 5 presents our within-country analysis for Mexico. Section 6 concludes. Additional robustness checks and detailsare provided in the online Appendix.

2. A MODEL OF MALTHUSIAN CONFLICT

In this section, we present a simple framework capturing the Malthusian notion thatpopulation growth may lead to social conflict (Malthus, 1798). The basic idea is thathigher population generates greater rents for fixed factors relative to labor, and thisscarcity makes conflict more likely. For less developed economies in 1940 or today, itmakes sense to think of land and natural resources as the scarce factor (henceforth, justland).

An important conclusion from this framework is that not every increase in populationleads to conflict. Only population growth that intensifies the scarcity of land intensifiesconflict. Conversely, population growth associated with improvements in technology orinstitutions will not promote greater conflict, and nor will population growth in economieswhere scarce factors are unimportant.

Suppose that aggregate output is given by a constant returns to scale (differentiable)production function where the factors of production are land (in fixed supply) denotedby Z and labor, N . In addition, the production function is indexed by technology, A:

Y = F (Z,N,A) ≡ f (N) . (2.1)

Here F (·) exhibits constant returns to scale in (Z,N) and f , which gives output as afunction of labor, holding technology, A, and land, Z, constant, has diminishing returns.Consequently, when N increases and A and Z remain constant, the fixed factor ofproduction, Z, will become more scarce, and so output per worker, f (N) /N and themarginal product of labor, f ′(N) will decline. On the other hand, when increases in Nare accompanied by improving technology, output per worker and wage need not decline.

We assume the following simple allocation of resources. Each individual i in societysupplies one unit of labor inelastically and also owns a fraction θi of land. For simplicity,we suppose that all markets are competitive, though this is not important for our analysis.With these assumptions, individual income and consumption (without conflict) are givenby

ci (N, θi) = f ′ (N) + θi [f (N)−Nf ′ (N)] . (2.2)

The key observation from equation (2.2) is that a marginal increase in an individual’sconsumption from an increase in his landholdings is larger when population is greater,i.e.,

∂2ci∂N∂θi

= −Nf ′′ (N) > 0.

Put differently, land shares matter more for consumption when population is greater.The intuition is straightforward: with higher N , land rents are more important relativeto wages due to the diminishing marginal product of labor. This is the fundamental force


underpinning our Malthusian model of conflict, which implies that the greater scarcityincreases conflict over the scarce factor.

To explore this channel in greater detail, imagine society consists of two groups, 1and 2. All members within a group are identical. To simplify the discussion suppose thatboth groups are of size N/2 and population growth leaves relative shares unchanged. Letus also assume that during conflict each group loses a fraction ρ of its labor income.

Group j has probability pj of winning the conflict and if it does win, it capturesa fraction λ−j of the land of the other group, where λ is an inverse measure of the“specificity of assets” to groups (or to individuals within a group). With probabilityp−j = 1−pj , group j loses the conflict and a fraction λj of its land. Also for simplicity, weignore any first-mover advantage, and there are no deaths from any conflict. Moreover, aswe elaborate further below, for now we do not allow any voluntary transfers or concessionsto avoid conflict. Finally, we assume that all agents are risk neutral. Then the expectedbenefits from conflict, πj (N, θ, λ, ρ) , for group j are given by

πj (N, θ, λ, ρ) = −ρf ′ (N) + [pjλ−jθ−j − p−jλjθj ] [f (N)−Nf ′ (N)] . (2.3)

The first term of this expression represents the loss of labor income. Because of thiscost, no group may wish to initiate conflict. If there is going to be any conflict, it will bethe group, here group j, for which

pjλ−jθ−j − p−jλjθj > 0

that has an incentive to initiate conflict.The same reasoning as in our discussion of equation (2.2) implies that for the group

at the margin of initiating conflict, we have:3

∂πj (N, θ, λ, ρ)

∂N> 0.

Therefore, an increase in population makes the group that is more likely to initiateconflict more “pro-conflict”. As noted before, this result does not apply when A increasesin tandem with N .4 This highlights that the Malthusian mechanism says nothing aboutincreases in population per se. Rather, the predictions are about the level of population forgiven A or for increases in population that are unusually large relative to the technologicaland other processes that tend to increase A. This observation underpins one of theimplications we investigate in our empirical work: population surges should have a biggerimpact on conflict in economies that are already under pressure from either slow growthor other economic problems.

This simple framework generates some intuitive comparative static results. Supposefor specificity that group j is at the margin of initiating conflict. Then, a greatershare of resources accruing to the weaker group (θj) makes conflict more likely. Lowerdisruption costs (lower ρ) and lower asset specificity of the weaker group (higher λ−j)also increase the likelihood of conflict.5 The point about asset specificity is linked to theimportance of natural resources and agriculture relative to human capital and industry.

3. In particular,∂πj(N,θ,λ,ρ)

∂N= −ρf ′′ (N) + [pjλ−jθ−j − p−jλjθj ] [−Nf ′′ (N)] > 0 since f ′′ < 0

and for the group at the margin of initiating conflict pjλ−jθ−j − p−jλjθj > 0.4. A change that increases f ′(N) and f(N)−Nf ′(N) by the same proportional amount would not

change incentives for conflict. More explicitly, starting from πj (N, θ, λ, ρ) = 0, a change that increasesboth f ′ (N) and f (N) −Nf ′ (N) by the same proportion will leave πj (N, θ, λ, ρ) = 0.

5. The relationship between inequality and conflict is non-monotonic. If θj = θ−j = 1/2, anincrease in θ−j increases both inequality and the likelihood of conflict. But if θ−j < θj , then an increasein θ−j reduces economic inequality but still makes conflict more likely.


A more industrial economy heavily relies on production processes—such as factoriesand long supply chains—that can be easily disrupted with violence, and skilled humancapital, which is typically harder to expropriate. This increases specificity and discouragesconflict (Acemoglu and Robinson, 2006). In contrast, the costs of conflict are smaller ineconomies relying on agriculture and natural resources. This reasoning highlights anotherimplication we investigate empirically: population growth should lead to conflict overscarce natural resources and not necessarily increase conflict unrelated to resources.

Finally, a central question that we have so far ignored is why more efficient waysof redistributing resources do not prevent conflict. A plausible explanation concernscommitment problems (Acemoglu and Robinson, 2001, 2006; Fearon, 1998, 2004; Powell,2006; Acemoglu et al, 2012). To see this, consider the same environment in a dynamicsetting, but in each period there is a probability q < 1 that either group can initiateconflict. Assume all agents have a discount factor β ∈ (0, 1). To simplify the discussion,assume as well that after conflict there is a permanent redistribution of resources andnever any social conflict again, and that only cash transfers (and no asset transfers) arefeasible. Suppose again that it is group j that is considering to initiate a conflict. In thiscontext, the benefits from conflict for group j are proportional to 1/ (1− β) because ofdiscounting. If the group is sufficiently patient (β is high enough), then cash transfers ina given period are not sufficient to offset this gain. But group −j cannot make a crediblepromise to make the cash transfers in the future once the window of opportunity forinitiating conflict disappears. In this setting, there will be conflicts along the equilibriumpath even though more efficient ways of dealing with conflict exist. In particular, fixβ ∈ (0, 1), then there exists q such that for all q < q, the Markov Perfect Equilibriumwill involve conflict. Also, there exists q < q, so that for all q < q, all Subgame PerfectEquilibria involve conflict (see Acemoglu and Robinson, 2006). It is straightforward to seethat in this dynamic extension with voluntary transfers it is still the Malthusian channelthat is responsible for conflict, and thus the comparative statics we have highlighted willapply—greater scarcity will make conflict more likely.

3. DATA

In this section, we present our cross-country and within-Mexico data and sources.

3.1. Cross-Country Data

In our baseline analysis, we measure conflict as the ratio of the number of years in conflictto total years in a period. We refer to these periods with the beginning year (for example,we refer to the average number of years in conflict between 1940 and 1950 as conflict in1940).6 This measure captures general conflict incidence rather than the precise timingof conflict. In our context, this is appealing because we are interested in a relativelylong-term phenomenon: increases in population over a period of several decades, and thepotential response in terms of greater social conflict. Datasets sometimes disagree on theexact year when a conflict began, but there are fewer differences regarding the incidenceof conflict within a decade.

Our baseline dataset is version 4 of the Correlates of War (henceforth COW) dataset(Sarkees and Wayman, 2010). In these data, a civil war is defined as a war fought

6. After 1940, some countries became independent, others lost their independence, fragmented,or experienced a major changes in borders. For each country, we check when the respective datasetsconsider the country as entering or leaving the state system, and adjust our measures accordingly.


within state borders, between government and non-government forces, where the centralgovernment is actively involved in military action, with effective resistance from bothsides and with at least 1,000 battle-related deaths during the war. This is a relativelyhigh threshold of violence for inclusion compared with other sources, as we explain below.The main advantage of COW is that it reports civil wars since 1816, and this long dataseries allows us to run a simple falsification test using pre-existing trends in conflict.When using COW, we assign the number of years with conflict to the reference dates asfollows: wars from 1940-1949 are assigned to 1940, wars from 1950-1959 are assigned to1950, and so on.

Our second database, covering dates since 1946, is the Uppsala Conflict DataProject, in conjunction with International Peace Research Institute (UCDP/PRIOArmed Conflict Dataset Version 4, Gleditsch et al, 2002). We assign the number of yearsin conflict to reference dates as follows: 1946-1949 to 1940, 1950-1959 to 1950, 1960-1969to 1960, etc. In the case of reference year 1940, we divide the number of years in war by 4(as the data only start in 1946); for other reference years we divide it by 10. This datasetincludes conflicts where at least one of the primary parties is the government of a state,and where the use of armed force results in at least 25 battle-related deaths per year.The dataset includes four types of conflicts, and we use the two categories for internalconflict (“internal armed conflict” and “internationalized internal armed conflict”).

Our third database is Fearon and Laitin’s (2003) coding of civil war. These datacover the period 1945-1999, and the criteria are broadly similar to those of COW, exceptthat anticolonial wars are coded as occurring within the empire in question (e.g., Algeriain the 1950s is assigned to France). As with the other datasets, we count the number ofyears that have any incidence of war, and use our usual rule for assignment to referencedates (1940 corresponds to 1945-1949, 1950 corresponds to 1950-1959, etc.).7

To examine effects on the intensity of conflict and as a further robustness check, weuse information on battle deaths from the Center for the Study of Civil War (CSCW)’sBattle Deaths Dataset (Lacina and Gleditsch, 2005). We use version 3, compatible withthe UCDP/PRIO dataset instead of the COW dataset, since the former has a lowerthreshold of battle-deaths for inclusion and includes more conflicts. This also allows usto more specifically check the robustness of our results to potential mechanical effects;for example, the detection and measurement of civil wars may increase simply becausethe population is larger and the number of potential deaths is higher. We rely on their“best estimate” of annual battle-related deaths (again we assign deaths to reference yearsusing the rule: 1940 = 1940-1949, 1950 = 1950-1960, etc.)

We have partial data for the 65 countries listed in Appendix Table A-1 (see “BaseSample”) and complete data from 1940 or earlier for 52 countries (51 when using COWsince Austria enters the COW state system in the 1950s). Unfortunately, there are noreliable historical data on causes of death for sub-Saharan Africa during the period underinvestigation.

We also consider a number of control variables in our robustness exercises, all ofwhich are described in Appendix Table A-1.

7. Conflicts are included if they: involved fighting between agents of (or claimants to) a stateand organized, nonstate groups who sought control of a government, region, or change in governmentpolicies; killed at least 1,000 in total, with a yearly average of at least 100; at least 100 were killed onboth sides (including civilians attacked by rebels).


3.2. Mexican Data

The unit of analysis for Mexico is a municipality, of which there are about 2,400,covering the entire country. Population growth and other municipal characteristics—age composition, school and university enrollment, literacy and share of immigrants—come from historical national censuses, accessed through the Mexican National StatisticalAgency (INEGI). We use population data from the Archivo Historico de Localidades(AHL), compiled by the INEGI, for localities (the level of territorial division belowmunicipalities). Municipality population is then computed as the sum of the populationsof the relevant localities. We adopt Sellars and Alix-Garcia’s (2018) corrections formissing localities in the AHL data.

We also use data from Matsuura and Willmott (2009) to code droughts in the1960s, which we define as the number of months for which precipitation is below the5th percentile of the long-run distribution (1900-2008) of monthly rain per municipality.

Our measure of malaria suitability is an index of potential malaria (plasmodiumvivax) transmission based on temperature suitability during an average year (Gethinget al, 2011). The index is available for 1 km by 1 km cells. Values range from zero(completely unsuitable for transmission) to one (maximum suitability). We average thisindex across the cells in the municipality.

To measure conflict at the municipality level, we use information on social conflictin the 1960s collected by Fergusson et al (2018) from Excelsior and El Universal, theonly two independent newspapers with national coverage. From hand-coded news storieson protests, strikes, demonstrations, riots and marches using the universe of articlespublished in the 1960s, we construct counts of violent and non-violent protests. Morespecifically, we used all news stories about protests published from 1960 to 1969 in thesetwo newspapers. We first identified any news story including any of the following keywordsin the title, description, or main text: Protestas (protests) and the n-gram “protest?”,Huelgas (strikes) and the n-gram “huelg?”, Manifestaciones (demonstrations) and then-gram “manifesta?”, Disturbios (riots) and the n-gram “Disturbio?, Marchas (marches)and the n-gram “March?”. We then identified stories in which the title or description isrelated to violence, conflict, arms, social disorder, or aggression. Specifically, we againuse keywords to code news stories as related to violent protests: “agita”, “desorden”,“violencia”, “violacion”, “armado”, “agresion” and “conflicto”. To measure non-violentprotests we simply counted all news stories about protests that do not have thesekeywords for violence. We also use keywords to identify protests related to resourcesand those that are unrelated to resources. All of our measures are expressed in per capitaterms relative to 100,000 people in 1940. Prior to the 1960s, the ruling party (PartidoRevolucionario Institutional, PRI) had a strong grip on the country and on the media,so there was little or no protest. In contrast, during the 1960s, some parts of the countrybecame much more prone to protest, including violent conflicts in some instances, andour data confirm this.8

In robustness checks we use data on historical conflict compiled by Ramos-Toro(2018).9 Additional variables used in our robustness exercises are described in AppendixTable A-1.

8. Specifically: the student movement became important; peasants defied the National PeasantUnion (CNC), the PRI’s arm in the rural sector; and opposition parties started to challenge the PRI insome parts of the country. The PRI stayed in power both nationally and locally with a combination offraud and repression (Bartra, 1985; Bellingeri, 2003; Bezdek, 1973; Herrera Calderon and Cedillo, 2012).

9. Fergusson et al (2018) show that these protest counts are strongly correlated with later electoralsupport for political opposition, once reliable voting data become available for Mexico in the 1980s.


4. CROSS-COUNTRY EVIDENCE

In this section we present our cross-country evidence. Our focus is on two stage leastsquares (2SLS) estimates exploiting differences in predicted mortality. Nevertheless,for comparison we start with descriptive statistics and Ordinary Least Squares (OLS)estimates.

4.1. Descriptive Statistics

Table 1 presents descriptive statistics (sample means and standard deviations) for ourbaseline cross-country sample. We present these summary statistics for our entire sample,and for subsamples of countries split by income and by whether they have experienceddeclines in predicted mortality above or below the median. Countries above medianpredicted mortality are those expected to experience greater “exogenous” declines inmortality and increases in population.

The first two rows show an increase in conflict in our sample between 1900 and 1940.However, columns 6 and 7 show that this increase is not correlated with the later changesin predicted mortality. The increase is larger in absolute terms for the “above-median”sample (0.051 versus 0.038), but since the group of countries with below-median changesin predicted mortality starts off at a lower baseline incidence (0.006 versus 0.047), therelative increase is marginally larger in the “below-median” group. The next six rowsof column 2 show a general trend, evident across all measures, of increasing conflictfrom the 1940s to the 1980s. Columns 3-5 show that this increase is concentrated amongmiddle-income and especially among poor countries. More importantly, comparing thechange in our conflict measures from 1940 to 1980 in columns 6 and 7, we observe thatcountries above median change in predicted mortality exhibit larger increases in conflictthan those below the median change. For instance, the average years in conflict (perdecade) according to the COW measure rose from 0.98 years in 1940 to 2.09 years in1980 for countries with above median change in predicted mortality, while it decreasedfrom 0.44 years to 0.25 years for those with below-median change. Table 1 further showsthat this increase is accounted for by conflicts over natural resources, with no evidencefor a differential change in non-resource conflicts. Our regression analysis confirms thesepatterns and establishes their robustness.

Finally, we also see a larger increase in population for the above-median groupbetween 1940 and 1980, and no differential changes in population between 1900 and 1940(see rows 14-16).

4.2. OLS Results

We begin with simple OLS regressions of conflict on population. In Table 2 we reportregressions of the form

cit = πxit + ζi + µt + Z′itβ + εit, (4.4)

where cit is a measure of conflict for country i and reference year t, and xit is logpopulation. ζi denotes a full set of country fixed effects while µt represents a full set ofyear dummies; we always include both to remove time-invariant country-specific factorsand global trends affecting population and conflict. Zit is a vector of other controls. For all


of our regressions, we calculate standard errors that are robust against heteroscedasticityand serial correlation at the country level (e.g., as in Wooldridge, 2002, p. 275). 10

In Table 2 and hereafter, we present two types of estimates: long differences (PanelsA and C in Table 2) and panel regressions (Panels B and D). The long-differencesspecifications use data only from 1940 (i.e., the 1940s, assigned to 1940) and 1980 (i.e.,conflict in the 1980s, assigned to 1980). In these specifications, equation (4.4) is equivalentto a regression of the change in conflict between the two dates on the change in logpopulation between the same two dates, which yields a particularly simple interpretation.

Panel regressions use data for intermediate years with one observation per decade(i.e., t = 1940, 1950, 1960, 1970, 1980), and are unbalanced because of data availability.In Section 4.6 we investigate how the response of conflict to population growth changedover time.

The OLS results in columns 1, 2 and 3 of Table 2 reveal that population is positivelycorrelated with conflict. The estimated coefficient for log population in the long-differenceregression in column 1 of Panel A, 0.323, implies that a 10% increase in population willbe associated with a 0.3 more years of conflict in the 1980s (relative to the 1940s).Alternatively, this estimate implies that the average change in log population in oursample of 0.676 will be associated with 2.18 more years of conflict.11

The magnitude of this coefficient is fairly stable across different conflict datasets,as seen in columns 2 and 3, which use the UCDP/PRIO and Fearon-Laitin datasetsrespectively. To check the robustness of our results across alternative measures of conflict,column 4 considers log(1+ battle deaths per initial population) as the dependent variable.The coefficient for population remains positive and significant at 10%, and implies similarquantitative effects. Panel B shows analogous results from estimating equation (4.4) usingpanel data.

One possible concern with the results in Panels A and B is that they might bydriven by age composition effects. In particular, rather than larger populations beingassociated with more conflict, it may be that younger populations are an importantcausal factor (Urdal, 2006). Panels C and D address this issue by including the share ofpopulation from 15 to 34 years of age as an additional independent variable (we lose eightcountries due to lack of data). Although the share of young people, which is endogenousto population growth, is a ‘bad control’ (a term from Angrist and Pischke, 2008), thisspecification is nonetheless useful for verifying whether there is a correlation betweenchange in population and conflict beyond what is accounted for by the presence of largeryoung cohorts. The results are consistent with Panels A and B—the coefficients andsignificance level for log population are similar. The point estimate for the share of youngpopulation is negative and sometimes marginally significant in the panel regressions ofPanel D. At least in this OLS specification, having more young people—once we controlfor log population—actually reduces conflict.12

10. One concern is that these standard errors may be downward biased due to a small number ofclusters. We also implemented the wild bootstrap procedure suggested by Cameron et al (2008). Theseresults are presented in columns 1 and 2 of Appendix Table A-2 and confirm that our conclusions arenot affected by the exact method of computing standard errors. This is consistent with Cameron et al(2008) who find very similar rejection rates for the cluster robust and wild bootstrap standard errors intheir Monte Carlo simulations with 50 clusters.

11. This is 0.676 multiplied by 0.323, and then multiplied by 10 (as our dependent variable is thefraction of the decade that the country is in conflict). Similarly, the 0.3 number in the previous sentenceis obtained as 0.323 × 0.10 × 10 ≈ 0.3.

12. In Appendix Table A-3 we use the share of the population aged 20 to 39 as an alternative andobtain very similar results.


These OLS estimates are not necessarily causal, and the true effect of populationon conflict might be larger or smaller than implied by these coefficients. For example, ifpopulation increases coincide with general improvements in economic conditions or theinstitutional environment, their OLS relationship with conflict will be biased downward.We next investigate this issue with our instrumental variables strategy.

4.3. International Epidemiological Transition

Our identification strategy relies on the international epidemiological transition, whichgenerated significant improvements in health conditions and life expectancy across awide range of countries. At the root of this transition were: major innovations in drugs(e.g., penicillin, other antibiotics) and associated effective treatments; chemicals (e.g.,DDT); and broader public health campaigns spearheaded by international agencies suchas the World Health Organization, which spread best practices around the world. Theseinterventions had greater impact on countries where infectious diseases were initially moreprevalent. Because this wave of innovations and campaigns was not caused by conditionsin the countries of our sample, this episode provides a potentially exogenous source ofvariation in health conditions and thus in population in both our cross-country datasetand within Mexico.

For the cross-country analysis, we use the predicted mortality instrument fromAcemoglu and Johnson (2007). This measures is the sum of each country’s initial (1940)mortality rate from infectious diseases until the moment of a global intervention (aninnovation or campaign). After a global intervention occurs, the mortality rate from thedisease in question declines to the lowest or “frontier” mortality rate in our sample. Forcountry i at time t, the instrument is defined as

M Iit =

∑d∈D

((1− Idt)Mdi40 + IdtMdFt), (4.5)

where: Mdi40 denotes mortality in 1940 (measured as number of deaths per 100individuals per annum) for country i, from disease d ∈ D; Idt is a dummy for interventionfor disease d that takes the value of 1 for all dates after the intervention; MdFt is mortalityfrom disease d at the health frontier of the world at time t (taken to be zero mortalityas in Acemoglu and Johnson, 2007); and D is the set of 13 diseases we use.13

Since Mdi40 is the pre-intervention mortality rate for disease d, and Idt = 1 after aglobal intervention, the variation in this variable comes from the interaction of baselinecross-country disease prevalence with global intervention dates for those specific diseases.Countries that had higher mortality from infectious disease in 1940 are expected toexperience larger increases in population after the intervention.14

We use this variable as an instrument for population. Specifically, we posit the first-stage relationship for country i at time t,

xit = ϕM Iit + ζi + µt + Z′itβ + uit, (4.6)

13. These 13 diseases are (roughly in descending order of importance): malaria, pneumonia,tuberculosis, influenza, cholera, typhoid, smallpox, whooping cough, measles (rubeola), dyphteria, scarletfever, plague and typhus. Acemoglu and Johnson (2007) discuss 15 diseases, but we do not have sufficientcoverage for dysentery, and yellow fever does not add any useful variation in our sample.

14. The predicted mortality instrument depends on the choice for dating global interventions. Analternative “global mortality instrument” employed in Acemoglu and Johnson (2007) is independent ofthe coding of global interventions, assuming instead that each country’s initial mortality rate decreasesat the pace of the global mortality rate for the disease in question. Our results are robust to using suchalternative, as shown in Table A-2.


where: xit is the logarithm of population at time t; M Iit the predicted mortality

instrument; ζi is a full set of country fixed effects; µt are year fixed effects; and Zit

again represents a vector of other controls.Acemoglu and Johnson (2007) show that changes in predicted mortality lead to

major improvements in life expectancy and also significantly increased births as morewomen survived to childbearing age. In countries such as India, Pakistan, Indonesia,Ecuador and El Salvador, where predicted mortality declined by a large amount, therewere big gains in life expectancy and consequently sizable increases in population.In contrast, in parts of Western Europe, Uruguay, Argentina, South Korea andAustralia, where predicted mortality did not decrease as much, life expectancy remainedcomparatively unchanged and there were smaller increases in population.

The exclusion restriction for our IV strategy, Cov(M Iit, εit) = 0, where εit is the error

term in the second-stage equation, requires that the unique channel for casual effects ofpredicted mortality on conflict is changes in population. We provide a series of falsificationexercises suggesting that predicted mortality is not correlated with various pre-existingdeterminants of economic growth and conflict. Another important concern is that changesin predicted mortality may affect conflict not just through population but via changes inoverall health conditions, age composition or incentives to acquire education. We providesome evidence against these specific channels below, but this concern needs to be bornein mind in interpreting our IV estimates.

4.4. First Stages, Reduced Forms and Falsification

Table 3 (and Figure 1) takes a first look at the first-stage relationship between populationand predicted mortality, as well as reduced-form regressions and falsification tests. Werun the following long-differences regression (equivalent to a panel regression with twodata points):

∆yit1,t0 = α+ ϕ∆M Ii1980,1940 + εit. (4.7)

where α is a constant and ∆yit1,t0 = yit1−yit0 is the change in our dependent variable forcountry i between reference dates t0 and t1. Similarly, ∆M I

i1980,1940 = Mi,1980 −Mi,1940

is the change in the predicted mortality instrument between 1940 and 1980.In columns 1 and 2 the dependent variable is the change in log population from 1940

to 1980. Column 1 includes all countries in the base sample, and column 2 focuses onlow and middle income countries. In both cases, we observe a strong negative first-stagerelationship between log population and predicted mortality (illustrated in Panels 1a and1b of Figure 1).15 The coefficient estimate of ϕ in column 1, −0.782, is significant at lessthan 1% and sizable. It implies that an improvement in predicted mortality of 0.469 per100 (or 469 per 100, 000, which is the mean improvement between 1940 and 1980 in ourbase sample) leads to an increase of roughly 0.37 in log population—approximately a45% increase in total population. The mean population in our sample in 1940 was about34.7 million, so this estimate translates into an increase in population of 12.8 million.The actual mean increase in population between 1940 and 1980 was about 23.5 million.This implies that changes in predicted mortality account for approximately half of theincrease in population between 1940 and 1980.

In columns 3 and 4 (and Panels 1c and 1d of Figure 1), we turn to the relationshipbetween predicted mortality and the fraction of each decade in conflict from 1940 to 1980.

15. In Tables A-8, A-11 and A-13 in the Appendix, we show the panel versions of these reducedforms, as well as their robustness to alternative samples and to the inclusion of differential trends.


For the base sample and for low- and middle-income countries, we see that countries witha larger decline in predicted mortality experienced a larger increase in years in conflict.Given the negative relationship between predicted mortality and population in columns 1and 2, this translates into a positive effect of population on conflict as our 2SLS estimatesbelow show.

If this causal interpretation—from the international epidemiological transition topopulation and via this channel to conflict—is correct, we should not see a correlationbetween changes in predicted mortality and pre-1940 changes in population or conflict.We explore this falsification exercise in columns 5-8 of Table 3 by estimating a variant ofequation (4.7) where t1 = 1940 and t0 = 1900. Reassuringly, we find no such pre-trends,and the coefficient estimates are insignificant and very small relative to our reduced formestimates. Predicted mortality explains changes in population and conflict after 1940,but not before 1940.

These results offer further confirmation that there were no preexisting trends inpopulation or conflicts related to changes in predicted mortality. This bolsters ourconfidence that the predicted mortality instrument is not capturing some other economicor political trends.

The scatter plots in Figure 1, which show these first stages, reduced forms andfalsification exercises, further illustrate that the patterns in Table 3 are not driven byoutliers. In each scatter plot we highlight the data point for Mexico to place our within-country case in the broader context of our cross-country analysis. Mexico lies around themiddle of the distribution for changes in predicted mortality and features relatively highpopulation growth. Interestingly, it experienced no change in conflict when measured asyears in conflict in the decade (in fact, it did not experience a year with civil war ineither 1940 or 1980). This reflects the high fatalities thresholds and the limited focus ofconflict measurement in international datasets, but it masks variation in social conflict inthe form of citizen protests, some of them quite violent. This limitation of cross-countrydata is yet another motivation for our within-country analysis.

4.5. 2SLS Results

Table 4 presents our main results, which are the 2SLS estimates of the effect of populationon conflict. Our second-stage regression is given by equation (4.4), where populationis instrumented by predicted mortality as in equation (4.6). We report long-differenceregressions for 1940 and 1980 in Panel A and panel regressions for 1940-1980 in Panel B.This table shows that the effect of population on conflict is positive and highly significantin almost all specifications.

In column 1, the dependent variable is the share of years in internal conflict perdecade, as measured by the COW dataset. The coefficient of interest, π, is estimated at0.617. This implies that a 10% increase in population leads to about 0.62 more years ofconflict in the 1980s relative to the 1940s, or the average change in log population inour sample, 0.676, has been associated with a 4.2 more years in conflict during the 1980s(again relative to the 1940s).

We find similar results in our panel regressions for 1940-80 presented in Panel B(π = 0.609, significant at the 1% level). For example, for El Salvador, which experiencedan increase in log population of 0.46 during this period (a change in population from 1.6to 4.6 million), the IV estimate of 0.617 implies an effect of roughly 2.8 more years inconflict (0.617× 0.46× 10).


Both in long-differences and in panel specifications, the IV estimates are larger thanthe OLS results reported in Table 2. There are three reasons for OLS and IV estimates todiffer in general. First, OLS estimates may be attenuated due to measurement error in thekey right-hand side variable, in this case population. Though a possibility, we do not thinkthis is a major concern in this case since population numbers are fairly accurate. Second,OLS and IV estimates capture different local average treatment effects. Third and mostimportantly, OLS estimates are affected by omitted variable biases, in particular, becausechanges in population are likely to be correlated with overall economic conditions andpotential changes in the institutional environment, both of which could impact conflict.The signs of these correlations are ex ante ambiguous. The fact that OLS estimatesare smaller than IV estimates suggests that these correlations are positive, leading todownward bias in OLS correlations.

Columns 2-4 investigate the robustness of the IV estimates. The dependent variablesin columns 2 and 3 are the years in internal conflict as a fraction of total years inthe reference date as measured by the UCDP/PRIO and Fearon and Laitin datasets,respectively. All the estimated coefficients are positive, and typically significant at lessthan the 1% level, with the exception of the UCDP/PRIO regressions in Panel B.16

Since conventional measures of civil war rely on meeting a battle death threshold,an increase in total population may mechanically increase the number of “detected” civilwars. We use battle deaths data to examine whether this may be driving our results.Column 4 considers log battle deaths for each reference date, divided by population in1940, to calculate cit. The coefficient on population is again positive and significant.17

4.6. Alternative Samples, Instruments, World War II and Timing

Table 5 presents estimates of our first- and second-stage relationships based on equations(4.4) and (4.6) for alternative samples and specifications. Panels A and B report long-difference specifications, and Panels C and D report panel regressions.

Column 1 includes all countries in our sample. In Panels A and C this columnreproduces our baseline estimates from column 1 of Table 4 to facilitate comparison.Columns 2 and 3 repeat the same set of regressions excluding Eastern and WesternEurope respectively. The estimates are broadly similar, suggesting that our results arenot unduly affected by the various conflicts before and during World War II, which wereconcentrated in Europe. In Table A-7 in the Appendix, we also show that the results arevery similar when we exclude Asia, Latin America, Australia or Africa from our sample.

Column 4 drops initially rich countries to verify that the results are not driven by thecomparison between rich and poor nations. Columns 5-7 check whether results are driven

16. We have a sparse distribution of countries around the globe, so strong spatial correlationbetween them is not a first-order concern, with the exception of India, Pakistan and Bangladesh, whichare treated as a single cluster. Moreover, our main results are robust to correcting the standard errorsfor spatial correction. In Appendix Table A-4 we report similar results when we apply Conley’s (1999,2010) correction for spatial correlation using a maximum radius of 9,684 km (the results are very similarwith other distances).

17. In columns 1 and 2 of Table A-5 in the Appendix we show that the results are similar if we usethe inverse hyperbolic sine transformation of the ratio of battle deaths to population rather than log 1+battle deaths. Columns 3 and 4 of the same table show that the results are also robust when we use adummy for any conflict in the decade.

Appendix Table A-6, on the other hand, shows that there appears to be no effect of populationgrowth on external wars or on other political outcomes, with the exception of state failure, which is itselflikely to reflect the effects of civil wars and other severe internal conflicts. This pattern suggests that themain effect of rapid population growth may be on civil wars, civil conflict and other internal tensions,but not on inter-state conflict and not even other aspects of political change in the country.


by events around World War II. Column 5 excludes the countries demographically mostaffected by that war, namely Austria, China, Finland, Germany, Italy and the RussianFederation (Urlanis, 2003). Column 6 assigns the level of conflict of the 1950s to the1940s. Column 7 drops the war years completely and assigns the number of years inconflict from 1946-49 (as a fraction of the 4 years in these interval) to our dependentvariable in 1940. The results are similar in all cases.

Columns 8 and 9 investigate whether our IV results are driven by the impact ofchanges in predicted mortality on age composition or the overall health of the population.Column 8 controls for the share of the young (15-34) in the population. Despite beingendogenous to changes in population, this variable is not statistically significant in ourregressions and the effects of population on conflict are very similar to our baselineestimates. Column 9 controls for (log) life expectancy at 20 as a measure of the overallhealth of the adult population. The ‘bad control’ concerns are more pronounced for thisvariable. Nevertheless, life expectancy at 20 is itself insignificant and the estimates of theimpact of population on conflict are similar to our baseline results, even if less precise.18

Finally, column 10 extends the analysis to the 2000s. There is again a significantimpact of population on conflict, though the size of the effect in the 2SLS estimate is about50% smaller than our baseline estimate (or 25% smaller in our panel specifications).19 Interms of the first-stage relationship, the estimates of ϕ in all of these cases are similarand significant at less than 1%.

Panels C and D have the same structure as in Panels A and B, now using a panelregression with decadal observations. The results are still highly significant and broadlysimilar. Table A-8 in the Appendix verifies that our reduced-form regressions are robustto the same specification checks.

4.7. Differential Trends

We next explore the robustness of the first- and second-stage relationships presented sofar to differential trends related to other factors. Specifically, we augment the first-stageequation with interactions between various baseline characteristics and year dummies, asindicated in the following equation:

xit = ϕM Iit + ζi + µt +

1980∑t=1940

κ′iωt + uit, (4.8)

where ωt = 1 in year t and zero otherwise, and κi are “time-invariant” characteristics ofcountry i. Similarly, we estimate the following second-stage equation:

cit = πxit + ζi + µt +

1980∑t=1940

κ′iωt + εit. (4.9)

18. The ‘bad control’ issue is somewhat less important for the share of the young populationbecause, as we show in Appendix Table A-9, predicted mortality does not have a major impact on thesize of young cohorts (those between 15 and 34 or between 20 and 39).

19. Appendix Table A-10 examines how the response of conflict to population growth changed overtime in more detail. Columns 1-5 in the two panels report long-differences and panel regressions wherethe initial time period is t = 1940 and the final date is 1960, 1970, 1980, 1990 and 2000. The effects areweaker when we end the sample in 1960 or in 1970. This might be because the impact of populationon conflict is delayed as it takes some time for the effects of population-induced scarcity to work out.The effects are also smaller when we extend the sample to the 1990s or 2000s. This is partly becauseof the end of the Cold War. Indeed, in column 6 in Panel B of Table A-10 we show that the effects ofpopulation on conflict are greater in countries where there was greater US or Soviet involvement duringthe Cold War (as measured by the number of CIA and KGB interventions coded by Berger et al, 2013).


These characteristics include: a measure of the average quality of institutions(average of the constraints on the executive from the Polity IV data set over 1950-70); initial (1930) GDP per capita and population; initial (1930) civil war; and latitudeand a malaria ecology index as geographic controls; share of young people (15-34)and share of adults (20-39); measures of natural resources;20 and ethnic polarization,21

which are often emphasized in the empirical literature on civil war.22 Table 6 showsthat our estimates of ϕ and π are similar across specifications both in long-differencesand the panel specifications. Only in one case, when we include time interactions withGDP per capita in 1930 in column 2, the second-stage estimate increases notably—forexample, from 0.617 in Table 4 to 1.132 in the long-differences specification. The effectsof additional baseline characteristics are explored in Table A-12 in the Appendix, againwith similar results.

Overall, the results in Table 6 suggest that our estimates so far are unlikely to beconfounded by differential trends related to baseline differences across countries.

4.8. Instrument Robustness Checks

As recommended in recent work by Goldsmith-Pinkham et al (2018), we investigated therelative importance of different diseases for our results. Specifically, Table A-14 reportsthe “Rotemberg weights” which decompose the Bartik estimator into a weighted sum ofthe just-identified instrumental variable estimators using each disease share as a separateinstrument.

Pneumonia, tuberculosis and malaria (and to a lesser extent, influenza, typhoid andmeasles) in the 1940s are the key drivers of our results in both the panel and long-difference specifications. This is in line with our prior based on mortality rates. TablesA-15 and A-16 report first stages and second-stage estimates with a number of differentspecifications and different disease compositions to verify that individual diseases are notresponsible for our results.

The estimates are similar when we exclude the diseases with the highest Rotembergweights, pneumonia, malaria and tuberculosis. When we exclude all three at the sametime in column 5, the first-stage coefficient changes significantly, but this is most likelydue to power issues and the 2SLS estimate in Table A-16 is similar. In columns 7-9 wereport estimates using just pneumonia, just malaria or just tuberculosis. In these casesthere are significant differences in the size of the 2SLS estimate (it is quite a bit smallerwith pneumonia than with just tuberculosis), but in all specifications we continue to

20. Sachs and Warner (1999) have popularized the share of natural resource exports in GDP. AsRoss (2006) notes, however, this measure may be a poor proxy of rents in the economy or potentialrevenues for the government since it does not include oil that is produced but consumed domestically,and it does not account for extraction costs which may vary across countries. Also, even at similar levelsof production, the numerator tends to be larger in poor countries because poor countries consume less oftheir own oil. Normalizing by GDP similarly inflates the numbers for poor countries. Motivated by thisreasoning, in columns 5 and 6, we focus instead on diamond production per capita (from Humphreys,2005) and oil and gas rents per capita (from Ross, 2006). We found similar results with oil productionper capita (also from Humphreys, 2005).

21. In the text we focus on an index of ethnic polarization constructed by Montalvo and Reynal-Querol (2005) based on the theoretical work by Esteban and Ray (1994). We present results with measuresof ethnic divisions in Appendix Tables A-18-A-21.

22. We include a malaria ecology index as an additional control given the emphasis that it hasreceived in some of the literature, which enables us to show that our results are not unduly affectedby malaria (see also the results in the next subsection exploiting variation in one disease at a time).In our within-Mexico analysis, we use within country variation in malaria conditions as a key source ofvariation in mortality and population.


estimate a statistically significant impact of population on conflict. The overall findingsare consistent regardless of which disease subsets we focus on.

4.9. Mean Reversion

In response to Acemoglu and Johnson’s (2007) finding of no effects of life expectancyon income, Bloom et al (2014) argued that the level of life expectancy in 1940 affectedsubsequent growth rates and should be included on the right-hand side. Acemoglu andJohnson (2014) showed that controlling for initial life expectancy or mean reversiondynamics does not change the finding of no positive impact of life expectancy on income.Although we suspect these potential dynamics to be less relevant for conflict than forincome, we explore their possible role in our results in Table 7. Our first strategy is toinclude the initial level of conflict or population interacted with time dummies in ourdecadal panel to flexibly control for mean reversion and the impact of past populationlevels. Our second strategy allows lagged conflict to linearly impact current conflict.Finally, we also present results from a nonlinear model that imposes the specification ofconvergence dynamics proposed by Bloom et al (2014).

For comparison, column 1 reports our baseline panel estimates (as in column 1 inTable 4, Panel B). Column 2 restricts the sample to countries with available informationon initial conflict, reducing the set of countries from 65 to 58. This has essentially no effecton our key point estimate, which changes from 0.609 to 0.606 and remains statisticallysignificant at 1%.

Column 3 includes the interaction of initial war with a full set of year dummies.The coefficient on population changes only slightly, to 0.584 with a standard error of0.181, even though the interactions between time dummies and initial conflict are jointlysignificant (the p-value, under 0.01, is reported at the bottom of the table). Columns4 and 5 add lagged conflict on the right-hand side, allowing for mean reversion inconflict. Column 4 uses the standard 2SLS estimator and column 5 presents Arellanoand Bond’s (1991) optimally weighted two-step generalized method of moments (GMM)estimator, with predicted mortality as the external instrument. This further reducesthe sample because it requires additional predetermined lags of conflict for estimation.Lagged conflict is not significant in either column, suggesting that mean reversion is nota major issue. Although the point estimate for population falls (to 0.266 with a standarderror of 0.107 in column 4 and 0.238 with a standard error of 0.105 in column 5), theresults are again broadly consistent with a positive and significant effect of populationon conflict.

Columns 6, 7 and 8 modify the specifications of columns 3, 4 and 5 to includeinteractions of initial population (rather than initial war) with a full set of time dummies.Our key point estimate in column 6 is very close to the baseline estimate of 0.609 andremains statistically significant at 1%. Controlling for lagged conflict has little impacton the standard 2SLS estimator (column 7) or on Arellano and Bond’s GMM estimator(column 8), and lagged conflict is not significant in either column. We thus conclude thatmean reversion appears unimportant and the effect of population remains positive andsignificant (although somewhat less precise in column 8).

Finally, for completeness in column 9 we consider the error correction modelproposed by Bloom et al (2014),

∆cit = αt + π∆xit + λπxi,t−1 − λci,t−1 + εit.


In this error correction model, conflict adjusts slowly to its steady state value. To estimatethis equation, we use the optimally weighted two-step GMM estimator with the secondand all longer lags of conflict and the first and all longer lags of predicted mortality asinstruments. We continue to estimate a positive and statistically significant (long-run)impact of log population on conflict, given by the parameter π (0.445 with a standarderror of 0.038). Moreover, the point estimate for λ is negative (−0.062 with a standarderror of 0.008), and confirms that there is little mean reversion in conflict. At the bottomof the table, we present the long-run impact of a 10% increase in population on conflictfrom the various different models. Though the short-run effects vary across columns,long-run impacts are quite similar across the various specifications. In particular, in theerror correction model estimated by GMM this long-run impact is 0.045 compared to0.061 in our baseline model.

4.10. Mechanisms

Our Malthusian model presented in Section 2 makes two additional testable predictions.First, population increases should lead to greater conflict related to natural resources (asopposed to non-resource related conflicts for which we should not see much of an effect).Second, it should be increases in population without corresponding improvements inproductivity and thus GDP that lead to greater conflict. We now explore these twopredictions. We follow Rustad and Binningsbø (2012) who have classified all internalarmed conflicts in UCDP/PRIO as related or unrelated to natural resources. In Table 8we use this classification to investigate the impact of population on resource and non-resource conflicts. Consistent with our expectations, column 1 shows a sizable impact ofpopulation on resource-related conflicts. Quantitatively, this estimate is larger than ourbaseline estimate in Table 4 (e.g., 0.918 for resource-related conflicts compared to ourbaseline estimate of 0.617 for all conflicts). In contrast, column 2 indicates that thereis no relationship between population and conflict unrelated to natural resources—theestimate is small and insignificant.

Columns 3 and 4 provide additional evidence consistent with this interpretationby splitting the sample between countries with population density above or below themedian in our sample. The effects of population on conflict are significantly larger amonghigh density countries where we expect competition for resources to be more intense.

Columns 5 and 6 turn to the second prediction from our theory. We code two newdependent variables. The first one, used in column 3, Conflictslow, is equal to zero if thechange in log GDP per capita from 1940 to 1980 is above the median in the sample andequal to the change in conflict from 1940 to 1980 otherwise. The second one, Conflictfastwhich is used in column 4, is defined analogously and is equal to zero when growth is belowthe median and equal to the change in conflict from 1940 to 1980 otherwise. These twovariables capture in a simple way the incidence of conflict in countries experiencing slowand fast growth. Consistent with our theoretical predictions, there is a strong effect onConflictslow, capturing the fact that among the slow-growing countries high population isassociated with a significantly higher incidence of conflict. In contrast, there is no effecton Conflictfast, which suggests that growing economies can absorb higher populationwithout creating scarcity and thus higher conflict.

Growth during our sample period is potentially endogenous to changes in population,however. Columns 7 and 8 deal with this problem by focusing on variation coming fromthe world prices of commodities that a country exports. We use export data from the UNComtrade dataset and data on world prices of 24 commodities since 1940 (from Grilli


and Yang, 1988, and Pfaffenzeller et al, 2007). Using these data we construct an index

of commodity price shocks for each country and year as Pit =∑J

j=1 ∆pjt × ωji, where∆pjt is the yearly change in the price of commodity j and ωji is the export per capitaof commodity j in country i in the 1960s. We identify times of economic hardship forcountry i as years in which Pit is negative. We then construct the analogues of Conflictslowand Conflictfast variables in columns 5 and 6, Conflicthigh hardship and Conflictlow hardship,based on whether the number of years with hardship is above or below the medianof its distribution during the relevant sample period. Column 7 shows a large andprecise impact of population on conflict when there are negative commodity price shocks(Conflicthigh hardship), while in column 8 there is no such impact on Conflictlow hardship.These results confirm that the effects of population on conflict are much more pronouncedin the presence of economic hardship.

As noted in the Introduction, changes in mortality may also induce greaterinvestments in education (Hansen and Strulik, 2017). These changes might in turnincrease violence because of unmet aspirations (Krueger, 2004; Campante and Chor,2012). To investigate this issue, in column 1 of Appendix Table A-17 we examine whethereducational attainment responds to population growth in our 2SLS specifications and findno evidence of a major impact of population on education. The remaining columns addeducational attainment as an additional determinant of conflict (using different measuresof conflict). Clearly, education is potentially a ‘bad control’, caused partly by changes inpopulation. The fact that it is not predicted by instrumented log population partiallyalleviates this concern, but the results should still be interpreted with caution. In any case,we do not find any consistent or significant effect of education on conflict (either in longdifferences or in the panel specification), and the impact of population continues to bepositive and significant (even if its magnitude is somewhat smaller in some specifications).Overall, we do not find any evidence that education is an important channel linkingchanges in mortality and conflict.

Finally, we explored the heterogeneity of the effects of population with respectto several other economic, political and social variables in Appendix Tables A-18 andA-19 by interacting population with various baseline characteristics (including thoseused in Table 6) and in Tables A-20 and A-21 by splitting the sample into countriesabove or below the median value of each characteristic.23 In summary, AppendixTables A-18 and A-19 present some evidence that countries with greater ethnolinguisticfragmentation respond more strongly to population pressures. The interactions with theother characteristics are not significant or robust. In no cases do these interactions alterthe main effects of population. When we split the sample, there are some differences inthe magnitudes of the population effects, but these differences are neither statisticallysignificant nor robust.

5. WITHIN-COUNTRY EVIDENCE: MEXICO

We next turn to an investigation of the within-country variation in Mexico. Our empiricalstrategy mimics the one we used in the cross country, except that instead of the 14 diseases

23. These characteristics are: GDP per capita, ethnic polarization, ethnic fragmentation, religiouspolarization, religious fragmentation, average ethno-linguistic fragmentation, ethnic dominance, ethnicinequality, oil inequality, country area, agricultural climatic suitability, agricultural soil suitability,average combined agricultural suitability (the product of the soil and climate components), cropsuitability index, agriculture as a percentage of GDP (value added), educational attainment, and initialinstitutions (constraints on the executive and democracy scores).


we used there, we focus on variation from malaria in the baseline period and then tracethe implications of major anti-malarial campaigns of the 1940s and 1950s.

5.1. Malaria in Mexico

Malaria was a major cause of death in Mexico in the first half of the 20th century,and in 1936 the fight against malaria was declared a matter of national interest. In linewith the broader international epidemiological transition, the 1940s and 1950s markedan intensification of these efforts.24

The introduction of DDT dramatically increased the effectiveness and reach of thesecampaigns (Blancarte-Melendez and de Jesus Cabrera-Palma, 1959). DDT was initiallyapplied in parts of the country where malaria was endemic. But by 1948 this approachhad been extended nationwide (Rabell et al, 1986), and regular use of DDT was “fullygeneralized” by the second half of the 1950s (Dıaz-Barriga et al, 2003; Centro Nacionalde Salud Ambiental, 2000).

Consequently, there were major impacts on morbidity and mortality within lessthan a decade (Bleakley, 2010; Pan-American Health Organization, n.a.). Malaria casesdecreased from 41,000 in 1955 to 4,000 in 1960 (Dıaz-Barriga et al, 2003). Malariaaccounted for about 8% of all deaths in Mexico in 1931. This fell to 5.2% in 1940.By 1960, the rate had declined dramatically, to 1.75% of total deaths. By 1970 malariaaccounted for a negligible 0.01% of total deaths in Mexico (INEGI, 2015).

These averages mask large geographical variation. Areas where malaria was endemicbenefited significantly more than others. Available state-level data for 1940 confirms this.Malaria deaths were particularly high in the states of Tabasco (24.1% of all deaths),Oaxaca (18.2%), Chiapas (18%), Campeche (14.7%), Guerrero (12.7%), Morelos (11.1%)and Veracruz (11%). In contrast, in Aguascalientes, Zacatecas, Guanajuato, Durango,Mexico, Coahuila, Baja California Sur, Tlaxcala, Chihuahua and the Distrito Federaldeaths from malaria accounted for no more than 1% of total deaths (INEGI, 1941). By1960, malaria deaths (as a fraction of all deaths) had come down sharply in the high-malaria states and stood at: 7.3% in Tabasco, 8.9% in Oaxaca, 4.8% in Chiapas, 0.8% inCampeche, 3.5% in Guerrero, 0.6% in Morelos and 4.7% in Veracruz.

5.2. Descriptive Statistics

Descriptive statistics for our Mexico sample are provided in Table 9, both for the entiresample of Mexican municipalities and separately for those with malaria suitability (orequivalently predicted mortality) above or below the median.

From 1960 to 1969 there were, on average, 2.3 violent protests per hundred thousandinitial (1940) inhabitants, but crucially this number is larger in areas that have above-median malaria suitability (3.0 versus 1.64). There are no similar differences for non-violent protests (12.04 in high malaria areas and 11.94 in low malaria areas).

The gap between high and low malaria suitability municipalities is most pronouncedfor protests related to natural resources (1.52 versus 0.54), while there is a smallercontrast for other protests (1.47 versus 1.1). There is no clear gap in historical conflicts(0.041 versus 0.055).

24. Dıaz-Barriga et al (2003) note that anti-malaria campaigns in Mexico followed the generalexpansion of DDT (Dichlorodiphenyltrichloroethane) usage as an insecticide, following the successfulapplication of DDT by the US military in the Pacific during WWII. More context on internationalcampaigns is provided by Alilio et al (2004) and Pampana (1954).


Finally, the table shows that population levels in 1940 are similar between areas withhigh and low malaria suitability, but subsequently, there is faster population growth inareas with high suitability to malaria: the average increase in log population from 1940to 1960 is 0.46 in municipalities with above median suitability and 0.36 in municipalitiesbelow the median. This is in line with our expectation that anti-malarial campaignsincreased life expectancy and births in parts of Mexico previously suffering from highmalaria incidence.

5.3. Empirical Strategy

We use a similar identification strategy as in our cross-country analysis, but focusingon the role of malaria. This is mainly because we do not have data on mortality fromdifferent diseases at the municipality level in Mexico, but recall that malaria was amajor source of differential changes in mortality and population during this time period.We construct a predicted mortality instrument, Mi,t, across municipalities in Mexico,

where in the base period (1940) Mi,t is equal to malaria suitability. With the same logicas our cross-country predicted mortality instrument, after the anti-malarial campaignsare underway, in particular in 1960, we set Mi,t = 0 in all municipalities. Therefore,our within-Mexico predicted mortality instrument uses only information from baselinedistribution of malaria suitability.

For our dependent variable, ci,t, we exploit the fact that in the 1940s there wereessentially no protests against the PRI, and we have reliable data on protests in the1960s (where ci,t is normalized by population as in our cross-country work). Thus, our2SLS model mimics the cross-country models, in particular the second-stage equation is(4.4) instrumented by (4.6), that is,

ci,t = πxi,t + ζi + µt + Z′i,tβ + εi,t, (5.10)

xi,t = ϕMi,t + ζi + µt + Z′i,tβ + ui,t, (5.11)

where the only difference from our cross-country equations is that the predicted mortalityvariable, Mi,t, is based on a single disease, malaria, and uses the index of malariasuitability (rather than baseline mortality rates). In addition, xi,t is log population inyear t, now for municipality i, and municipality fixed effects and time fixed effects aredefined as in our cross-country equations. The vector Zi,t again designates the covariatesand now includes interactions between dummies for Mexican states and the post anti-malarial campaign dummy. These interactions partially control for the substantial cross-state differences within Mexico. Since spatial spillovers and correlation are a greaterconcern for smaller geographic units, in our within-Mexico work we always allow forspatial correlation between municipalities (Conley, 1999, 2010).25

5.4. First Stages, Reduced Forms and Falsification

We start by estimating the first-stage relationship summarized in equation (5.11) incolumn 1 of Table 10.26 Panel A of the table focuses on bivariate relationships without any

25. We use a radius of 35.9 km for the distance cutoff. This corresponds to the 80th percentile ofthe distribution of the average distance between a municipality and its closest neighbor, and thus ensuresthat the error terms for most municipalities are allowed to be correlated at least with their first-orderneighbors.

26. As in equation (4.7) in the previous section, we report these in the form of long differences,which are algebraically equivalent to equation (5.11) with two data points.


controls (except state fixed effects). Column 1 shows that the within-Mexico predictedmortality variable, based on malaria suitability, is a strong predictor of municipalitypopulation growth. The coefficient of Mi,t is −0.287 with a standard error of 0.042. Thisimplies that moving from a municipality that had no suitability for malaria in the baseperiod to one with average suitability (0.43) increases log population by roughly 0.12(0.43×−0.29), or equivalently, it raises population by about 12%. Column 2 presents thereduced-form between violent protests and our predicted mortality measure, revealing astrong and significant relationship (the coefficient is −3.52 with a standard error of 1.24).In this case, the move from no suitability to average suitability is associated with 1.5additional protests. Figure 2 depicts the first-stage and reduced-form relationships.27

In Panel B, we verify the robustness of these estimates to additional controls,which include basic geographic controls, initial population, measures of initial education,estimates of historical conflict (except in column 9), and measures resource abundanceand ethnic composition (these controls are described in detail in the next subsection inthe context of our discussion of Table 11). Both the first stage and the reduced formremain strong and statistically significant, even if the magnitudes of the coefficientsdecline somewhat when we add these controls (to 0.21 for the first stage and to 2.75 forthe reduced form).

Columns 3-7 of Table 10 conduct a number of falsification exercises. We first explorewhether our instrument also predicts population changes before 1940, specifically from1930-1940 (column 3) or from 1920 to 1940 (column 4).28 A significant correlation in thesespecifications would contradict our underlying assumption that the impact on populationafter 1940 comes from the anti-malarial campaigns rather than from other differentialtrends affecting municipalities with different levels of initial malaria. Reassuringly, ourpredicted mortality instrument is not robustly correlated with population growth in thepreceding 10-year or 20-year period, and the point estimates are much smaller than incolumn 1. Without controls (Panel A), the coefficient on our within-Mexico predictedmortality variable is −0.043 with a standard error of 0.024 in column 3 and −0.065with a standard error of 0.046 in column 4. When controls are included in Panel B, thecoefficient is further reduced to −0.001 in column 3 and to 0.02 in column 4.

We also show in columns 5 and 6 of Table 10 that the change in predicted mortalityis not correlated with previous changes (1930-1940) in either the share of the young adultpopulation (ages 20 to 39) or literacy. Finally, in column 7 we explore whether historicalviolence and social conflict measures are higher in municipalities with different patternsof predicted mortality. For this exercise, we rely on information on historical battles inMexico (from 1616 to 1940) digitized and geolocated by Ramos-Toro (2018) based on

27. Appendix Figure A-1 shows the geographical variation in malaria suitability (left panel),population growth (middle panel) and violent protests (right panel). We observe wide variation in malariasuitability across the country. In line with Figure 2, this variation is correlated with areas that experiencedmore rapid population growth and more intense protest activity following the anti-malaria campaign.

Appendix Table A-22 presents some back of the envelope calculations suggesting that themagnitude of the first-stage relationship is plausible. In particular, if we incorporate the positive effectson births from lives saved from malaria (documented in Acemoglu and Johnson, 2007), assume the samerelationship between comorbidity from malaria and other diseases in Mexico as in our cross-countrysample, and suppose that municipality-level malaria deaths declined linearly to the national average afterthe anti-malaria campaign, we estimate that a municipality with average suitability to malaria shouldhave experienced a 9.41% increase in population between 1940 and 1960 relative to a municipality withno malaria suitability. This is a little smaller than but in the same ballpark as our estimated 12.47%increase in population.

28. Because of the turmoil following the Mexican Civil War, the 1920 census was conducted in late1921 and for this reason may be less reliable. This motivates our greater emphasis on data from the 1930census for the falsification exercises.


work by Clodfelter (2002). Bolstering confidence in our empirical strategy, the change inpredicted mortality is not correlated with previous incidence of historical violence.

5.5. Two-Stage Least Squares Results and Robustness

In this subsection, we present our main 2SLS results for the impact of population onviolent protests across Mexican municipalities between 1940 and 1960. Table 11 presentsthe 2SLS (Panel A), first-stage (Panel B) and OLS (Panel C) estimates.

Column 1 presents estimates of equation (5.10) without any additional interactionsor controls. Both the OLS and IV estimates are positive and significant at less than 1%.Our more limited data for Mexico prevent us from controlling for the same set of baselinecharacteristics we explored in the cross-country analysis. Nevertheless, we have data onbasic geographic features, initial population, measures of initial education, estimates ofhistorical conflict and proxies for resource abundance and ethnic diversity. In column 2(and all subsequent columns) of Table 11 we include the following geographic controls:distance to Mexico City, distance to the nearest big city,29 and a land quality index.In column 3 we also control for initial (1940) population. Column 4 controls for theinitial level of education, using the primary school enrollment rate in 1940 (the resultsare similar when we use literacy rates for the entire population or for young cohorts).Column 5 uses the university enrollment rate as an alternative measure of education.Column 6 controls for the historical level of conflict. Column 7 controls for the shareof a municipality’s area on a sedimentary basin, which is associated with the presenceof petroleum reserves and is unrelated to endogenous exploration efforts (see Cassidy,2018). Column 8 controls for the share of people who only speak indigenous languagesin the 1940 census. Finally, in column 9 we include all of these controls simultaneously.

The coefficient on log population is statistically significant across all specifications,and its magnitude is relatively stable, ranging from 10.9 to 14.3 in the 2SLS specifications(Panel A). Moreover, the first stages reported in Panel B are robust across specifications,and the coefficient of the predicted mortality variable ranges from −0.28 to −0.21 and isalways significant at less than 1%. Quantitatively, the 2SLS estimate in column 9 (13.05)implies that a 10% increase in population is associated with 1.3 more protests per hundredthousand inhabitants. Alternatively, the average change in log population in our sampleof 0.41 is associated with 5.35 more protests. As in our cross-country estimates, the OLScoefficients in Panel C are smaller than the IV estimates. Our interpretation is againthat the OLS estimates are biased downwards because of potential correlation betweenpopulation changes and changes in other economic and institutional factors reducingconflict.

5.6. Mechanisms

We next investigate the same two implications from our conceptual framework whichwe explored in the context of cross-country data. These center on whether populationincreases lead to conflicts that are related to natural resources and whether these effectsare exacerbated when there is economic hardship.

Columns 1-4 of Table 12 explore whether population increases are associated withresource-related protests. We use keywords to classify violent protests into those related

29. This is defined as the linear distance from a municipality’s centroid to the nearest municipalitywith a total population of at least 100,000 in 1960.


to natural resources, such as water, land, mining and agricultural products. Violentprotests unrelated to any of these categories are coded as non-resource conflicts. Panel Aof the table presents 2SLS estimates, while Panel B shows the corresponding first stagesfor log population (the first stages for the interactions in columns 7-10 are presentedin Appendix Table A-23). Column 1 focuses on resource-related (violent) protests andincludes no controls other than state fixed effects, while column 2 adds our full set ofcontrols as in column 9 of Table 11. The corresponding regressions for non-resourceconflicts are in columns 3 and 4. The estimates show a robust effect of population onresource-related protests in columns 1 and 2. In contrast, the impact of population on non-resource protests is smaller and ceases to be statistically significant once our controls areadded. The point estimates in columns 2 and 4, for example, imply that a 10% increasein population leads to a 0.87 more resource-related protests per 100,000 municipalityinhabitants, while the same increase leads to only 0.43 more non-resource conflicts per100,000 inhabitants.

Columns 5 and 6 turn to nonviolent protests as the dependent variable. There is asignificant positive impact in column 5, but this is not robust to the same controls usedin column 6. Overall, the evidence suggests that violent protests are more responsiveto population pressures, though there appears to be some weaker effects on nonviolentprotests as well.

Columns 7-10 investigate the second implication of our conceptual framework andinvestigate whether the effects of population pressures are greater when there are othersources of economic hardship. In particular, we exploit the random occurrences of droughtacross municipalities. Droughts are defined as precipitation levels below the 5th percentileof the long-run distribution (1900-2008) of monthly rain per municipality. As noted in Dell(2012), drought is harmful during most of the year because it lowers soil moisture contentand reduces plant growth, but it is beneficial during the harvest season. We thereforefocus on the number of months with droughts in the 1960s occurring during the non-harvesting period for corn. If a Malthusian mechanism links population and conflict asin our conceptual framework, then municipalities facing greater economic hardships dueto droughts should experience greater increases in conflicts following population rises.Our estimates show that this is indeed the case and the interaction between droughtand population is positive and significant, and is robust to the inclusion of our standardcontrols. We also include interactions for droughts in the harvest season as a placeboto investigate whether any other mechanisms aside from resources might be connectingdroughts and violence.30 These results are reassuring: the coefficient for this interaction isnegative and not statistically significant.31 The models in columns 9 and 10 additionallyinclude interactions between the harvest and the non-harvest dummies and the post-anti-malarial comparing dummy as a control for any direct effects from droughts. Thisincreases the standard errors of our drought interactions, but the overall pattern remainssimilar.

30. Hsiang et al (2013) argue that warmer temperatures and extreme rainfall influence conflict.Scheffran et al (2012) discuss multiple possible causal pathways between weather and violence.

31. Droughts are demeaned when included in the interaction terms to facilitate interpretation ofthe direct effect of change in log population (which ensures that these effects are evaluated at themean of the drought distribution). In Appendix Table A-25 we verify the robustness of the results toalternative definitions of drought. For example, they are very similar when we define droughts usingdifferent percentiles of the long-run distribution of monthly rain. We also confirm that our results arerobust to using the drought severity measure of Dell (2012), which is the maximum between 1 and theratio of the decade’s mean monthly precipitation to long-run average monthly precipitation.


Overall, consistent with our conceptual framework and the results reported our cross-country analysis, it appears that increases in population caused a spike in violent protestsacross Mexican municipalities, and this impact was more pronounced in the presence ofdrought. Moreover, only droughts that affect yields (non-harvest droughts) raised theresponsiveness of violent protests to population pressures.

As in our cross-country analysis, a final question is whether changes due to lowermorbidity and mortality—resulting from the reduced incidence of malaria—are impactingconflict via changes in the share of younger cohorts in the population or through changesin education as opposed to our Malthusian mechanism.32 To examine these issues, PanelA of Appendix Table A-24 explores whether predicted mortality (Mexico) is correlatedwith changes in population shares or educational attainment. In Panel B, we use theshare of the population aged 20-39 and educational attainment variables as controls ina regression for violent protests on (instrumented) log population. The population shareand education measures are again likely ‘bad controls’ because they are affected by overallchanges in population, and in fact, in this case they do seem to be correlated with ourinstrument. All the same, our main effects remain largely unchanged when these variablesare included.

6. CONCLUSIONS

The large population increases in many developing countries that followed theinternational epidemiological transition of the 1940s contributed to an increase in internalviolent conflicts, including civil wars and violent protests. This pattern is present andhighly robust both across countries and across Mexican municipalities.

Our conceptual framework proposes an explanation for these patterns based ona Malthusian channel—a larger population increases resource scarcity and encouragesfighting over scarce resources. This conceptual approach highlights that the problemis not so much that higher population will always lead to greater conflict, but thatpopulation surges unaccompanied by corresponding increases in productivity or physicaland human capital investments will do so. We bolster this interpretation by showing thatour results are accounted for by increases in conflicts related to natural resources andthe effects are significantly larger in economies experiencing other sources of economichardship (slower growth or droughts).

Obvious questions for future research in this area include extending our analysis toother periods and samples, and investigating what economic changes and policy actionscan better accommodate upcoming population surges around the developing world andavoid greater conflict.

Acknowledgments. We thank the editor Nicola Gennaioli and four anonymous refer-ees as well as seminar participants at MIT’s development lunch, the University of Chicago,the XXII Annual Conference of the European Society for Population Economics at UCL,2017 AEA Annual Meetings, Universidad de los Andes and Universidad del Rosario.Ioannis Tokatlidis and Juan Camilo Yamin provided superb research assistance. We alsothank Horacio Larreguy for comments and Jennifer Alix-Garcia, Horacio Larreguy, JuanFelipe Riano, Diego Ramos-Toro and Emily Sellars for sharing their data.

32. We do not have data on life expectancy across municipalities to proxy for the general health ofthe population.


Supplementary DataSupplementary data are available at Review of Economic Studies online.

REFERENCES

Acemoglu, D., Egorov, G., and Sonin, K. (2012). Dynamics and Stability of Constitutions,Coalitions and Clubs. American Economic Review , 102, 1446-1176.

Acemoglu, D. and Johnson, S. (2007). Disease and development: The effect of lifeexpectancy on economic growth. Journal of Political Economy , 115, 925-985.

Acemoglu, D. and Johnson, S. (2014). Disease and Development: A Reply to Bloom,Canning and Fink. Journal of Political Economy , 122, 1367-1375.

Acemoglu, D. and Robinson, J. A. (2001). A theory of political transitions. AmericanEconomic Review , 91, 938-963.

Acemoglu, D. and Robinson, J. A. (2006). Economic origins of dictatorship anddemocracy. New York, NY: Cambridge University Press.

Alilio, M. S., Bygbjerg, I. C., and Breman, J. G. (2004). Are multilateral malaria researchand control programs the most successful? Lessons from the past 100 years in Africa.The American journal of tropical medicine and hygiene, 71, 268-278.

Angrist, J. D. and Pischke, J.S. (2008). Mostly Harmless Econometrics: An Empiricist’sCompanion. Princeton: Princeton University Press.

Arellano, M. and Bond, S.R. (1991). Some Specification Tests for Panel Data: MonteCarlo Evidence and an Application to Employment Equations. Review of EconomicStudies, 58, 277-298.

Ashraf, N., Fink, G., and Weil, D. N. (2010). Evaluating the effects of large scale healthinterventions in developing countries: The Zambian Malaria Initiative NBER WorkingPaper , 16069.

Barofsky, J., Anekwe, T. D., and Chase, C. (2015). Malaria eradication and economicoutcomes in sub-Saharan Africa: evidence from Uganda. Journal of health economics,44, 118-136.

Barlow, R. (1967). The economic effects of malaria eradication. American EconomicReview , 57, 130-148.

Bartra, A. (1985). Los herederos de Zapata: Movimientos campesinos posrevolucionariosen Mexico, 1920-1980. Mexico: Editorial ERA.

Bellingeri, M. (2003). Del agrarismo armado a la guerra de los pobres, 1940-1974 Mexico:El Colegio de Mexico, Centro de Estudios Internacionales.

Berger, D., Easterly, W., Nunn, N., and Satyanath, S. (2013). Commercial Imperialism?Political Influence and Trade During the Cold War. American Economic Review , 103,863-896.

Besley, T. J., & Persson, T. (2008). The incidence of civil war: Theory and evidenceNBER Working Paper , 14585.

Bezdek, R. (1973). Electoral oppositions in Mexico: Emergence, suppression, and impacton political processes. PhD Dissertation, Ohio State University.

Blancarte-Melendez, J. R., and de Jesus Cabrera-Palma, N. (1959). El programa deerradicacion del paludismo en Mexico. Salud Publica de Mexico, 1, 9-50.

Blattman, C. and Miguel, E. (2010). Civil war. Journal of Economic Literature, 48,3-57.

Bleakley, H. (2010). Malaria eradication in the Americas: A retrospective analysis ofchildhood exposure American Economic Journal: Applied Economics, 2, 1-45.


Bloom, D. E., Canning, D., and Fink, G. (2014). Disease and Development Revisited.Journal of Political Economy , 122, 1355-1366.

Bruckner, M. (2010). Population size and civil conflict risk: Is there a causal link?Economic Journal , 120, 535-550.

Cameron, A. C., Gelbach, J. B., and Miller, D. L. (2008). Bootstrap-based improvementsfor inference with clustered errors. The Review of Economics and Statistics, 90, 414-427.

Campante, F. R. and Chor, D. (2012). Why was the Arab world poised for revolution?Schooling, economic opportunities and the Arab Spring. Journal of EconomicPerspectives, 26, 167-88.

Case, A., and Paxson, C. (2009). Early life health and cognitive function in old age.American Economic Review , 99, 104-09.

Cassidy, T. (2018). The Long Run Effects of Oil Wealth on Development: Evidence fromPetroleum Geology. Mimeo. University of Alabama.

Centro Nacional de Salud Ambiental (2000). Situacion actual de la malaria y uso delDDT en Mexico.. Centro de Vigilancia Epidemiologica: Secretarıa de Salud.

Clodfelter, M. (2002). Warfare and Armed Conflicts: A Statistical Reference, 1500-2000..Jefferson: McFarland.

Colella, F., Lalive, R., Sakalli, S.O., and Thoenig, M. (2018). Inference with ArbitraryClustering Mimeo. University of Lausanne.

Collier, P. and Hoeffler, A. (1998). On economic causes of civil war. Oxford EconomicPapers, 50, 563-573.

Collier, P. and Hoeffler, A. (2004). Greed and grievance in civil war. Oxford EconomicPapers, 56, 563-595.

Conley, T. G. (1999). GMM estimation with cross sectional dependence. Journal ofeconometrics, 92, 1-45.

Conley, T. G. (2010). GMM estimation with cross sectional dependence. InMicroeconometrics (pp. 303-313). Palgrave Macmillan, London.

Dell, M. (2012). Path dependence in development: Evidence from the MexicanRevolution. Harvard University.

Dıaz-Barriga, F., Borja-Aburto, V., Waliszewski, S. and Yanez, L. (2003). DDT inMexico.. In Persistent Organic Pollutants. Berlin: Springer, 371-388.

Diehl, P. F., and Gleditsch, N. P. (2001). Environmental conflict. Boulder, CO: WestviewPress.

Easterly, W., and Levine, R. (1997). Africa’s growth tragedy: Policies and ethnicdivisions. Quarterly Journal of Economics, 112, 1203-1250.

Esteban, J. and Ray, D. (1994). On the Measurement of Polarization. Econometrica, 62,819-851.

Esteban, J. and Ray, D. (1999). Conflict and distribution. Journal of Economic Theory ,87, 379-415.

FAO/IIASA. (2010). Global Agro-ecological Zones (GAEZ v3.0). FAO, Rome, Italy andIIASA, Laxenburg, Austria.

Fearon, J. D. (1998). Commitment problems and the spread of ethnic conflict. In D. Lake& D. Rothchild (Eds.), The international spread of ethnic conflict: Fear, diffusion andescalation. Princeton: Princeton University Press.

Fearon, J. D. (2004). Why do some civil wars last so much longer than others? Journalof Peace Research, 41, 275-301.

Fearon, J. D. (2006). Ethnic mobilization and ethnic violence. In B. R. Weingast & D.A. Wittman (Eds.), The Oxford Handbook of Political Economy (pp. 852–68). Oxfordand New York: Oxford University Press.


Fearon, J. D., and Laitin, D. D. (2003). Ethnicity, insurgency and civil war. The AmericanPolitical Science Review , 97, 75-90.

Fergusson, L., Larreguy, H., & Riano, J. F. (2018). Political competition and statecapacity: Evidence from a land allocation program in Mexico. Lacea Working Paper ,No. 2018-0011.

Fluckiger, M., & Ludwig, M. (2017). Malaria suitability, urbanization and persistence:Evidence from China over more than 2000 years. European Economic Review , 92,146-160.

Fugro Robertson, Ltd. (2013). Fugro Tellus Sedimentary Basins of the World Maphttp://www.datapages.com/gis-map-publishing-program/gis-open-files/

global-framework/robertson-tellus-sedimentary-basins-of-the-world-map.Garfinkel, M. R., and Skaperdas, S. (2007). Economics of conflict: An overview. In

K. Hartley & T. Sandler (Eds.), Handbook of defense economics (Vol. 2, p. 649-709).Elsevier.

Gething, P. W., Van Boeckel, T. P., Smith, D. L., Guerra, C. A., Patil, A. P., Snow, R. W.,& Hay, S. I. (2011). Modelling the global constraints of temperature on transmissionof Plasmodium falciparum and P. vivax. Parasites & Vectors, 4.

Goldsmith-Pinkham, P., Sorkin, I., and Swift, H. (2018). Bartik Instruments: What,When, Why, and How. NBER Working Paper , w24408.

Gleditsch, N. P., Wallensteen, P., Eriksson, M., Sollenberg, M., and Strand, H. (2002).Armed conflict 1946-2001: A new dataset. Journal of Peace Research, 39, 615-637.

Grilli, E. R., and Yang, M. C. (1988). Primary commodity prices, manufactured goodsprices, and the terms of trade of developing countries: what the long run shows. TheWorld Bank Economic Review , 2, 1-47.

Hansen, C. W., and Strulik, H. (2017). Life expectancy and education: evidence fromthe cardiovascular revolution. Journal of Economic Growth, 22, 421-450.

Hauge, W., and Ellingsen, T. (1998). Beyond environmental scarcity: Causal pathwaysto conflict. Journal of Peace Research, 35, 299-317.

Herrera Calderon, F. & Cedillo, A. (2012). Challenging authoritarianism in Mexico:Revolutionary struggles and the dirty war, 1964-1982. New York: Routledge.

Homer-Dixon, T. F. (1991). On the threshold: Environmental changes as causes of acuteconflict. International Security , 2, 76-116.

Homer-Dixon, T. F. (1999). Environment, scarcity and violence. Princeton, NJ: PrincetonUniversity Press.

Hong, S. C. (2007). The burden of early exposure to malaria in the united states,1850 − 1860: Malnutrition and immune disorders. The Journal of Economic History ,67, 1001-1035.

Hsiang, S. M., Burke, M., and Miguel, E. (2013). Quantifying the influence of climateon human conflict. Science, 341.

Humphreys, M. (2005). Natural resources, conflict and conflict resolution. Journal ofConflict Resolution, 49, 508-537.

INEGI. (1941). Anuario estadıstico de los Estados Unidos Mexicanos 1941. Mexico:Instituto Nacional de Estadıstica y Geografıa.

INEGI. (2015). Estadısticas Historicas de Mexico 2014: 4. Salud. Mexico: InstitutoNacional de Estadıstica y Geografıa.

Kaplan, R. (1994). The coming anarchy. The Atlantic (February).Krueger, Alan (2004). What Makes a Terrorist: Economics and the Roots of Terrorism.

Princeton: Princeton University Press.Lacina, B., and Gleditsch, N. P. (2005). Monitoring trends in global combat: A new

dataset of battle deaths. European Journal of Population, 21, 145-166.

http://www.datapages.com/gis-map-publishing-program/gis-open-files/global-framework/robertson-tellus-sedimentary-basins-of-the-world-map

http://www.datapages.com/gis-map-publishing-program/gis-open-files/global-framework/robertson-tellus-sedimentary-basins-of-the-world-map


La Porta, R., Silanes, F. Lopez-de, Shleifer, A., & Vishny, R. (1999). The quality ofgovernment. Journal of Law, Economics and Organization, 15, 222-279.

Lucas, A. M. (2010). Malaria eradication and educational attainment: evidence fromParaguay and Sri Lanka. American Economic Journal: Applied Economics, 2, 46-71.

Maddison, A. (2003). The World Economy: Historical Statistics. Paris, France:Development Centre of the Organisation for Economic Co-operation and Development.

Maddison, A. (2006). The World Economy. Paris, France: Development Centre of theOrganisation for Economic Co-operation and Development.

Malthus, T. R. (1798). An essay on the principle of population, as it affects thefuture improvement of society: With remarks on the speculations of Mr. Godwin, M.Condorcet and other writers. London: Printed for J. John.

Matsuura, K. and Willmott, C. (2009). University of Delaware global surface airtemperature and precipitation climatology and monthly grids, 1900-2008. http:

//research.jisao.washington.edu/data/ud/.Michalopoulos, S. (2012). The Origins of Ethnolinguistic Diversity. American Economic

Review , 102, 1508-1539.Miguel, E., Satyanath, S., and Sergenti, E. (2004). Economic shocks and civil conflict:

An instrumental variables approach. Journal of Political Economy , 112, 725-753.Montalvo, J. G., and Reynal-Querol, M. (2005). Ethnic polarization, potential conflict

and civil wars. American Economic Review , 95, 796-816.Mueller, H. (2016). Growth and violence: argument for a per capita measure of civil war.

Economica, 83, 473-497Pan-American Health Organization. (n.a.). Antologıa de la Atencion a la Salud en

Mexico,1902-2002. https://www.paho.org/mex/index.php?option=com_content&

view=article&id=204:antologia-atencion-salud-mexico&Itemid=315

Pfaffenzeller, S., Newbold, P., & Rayner, A. (2007). A short note on updating the Grilliand Yang commodity price index. The World Bank Economic Review , 21, 151-163.

Posner, D. N. (2004a). Measuring ethnic fractionalization in Africa. American Journalof Political Science, 48, 849-863.

Posner, D. N. (2004b). The political salience of cultural difference: Why Chewas andTumbukas are allies in Zambia and adversaries in Malawi. The American PoliticalScience Review , 98, 529-545.

Pampana, E. J. (1954). Changing strategy in malaria control. Bulletin of the WorldHealth Organization, 11, 513.

Pampana, E.J., & Russell, P.F. (1955). Malaria: A World Problem. World HealthOrganization Chronicle, 31, 31.

Powell, R. (2006). War as a commitment problem. International Organization, 60, 169-203.

Rabell, C. A., Mier, M. and Rocha, T. (1986). El descenso de la mortalidad en Mexicode 1940 a 1980. Estudios demograficos y urbanos, 1, 39-72.

Ramankutty, N. , Foley, J. A., Norman, J. and McSweeney, K. (2002). The globaldistribution of cultivable lands: current patterns and sensitivity to possible climatechange. Global Ecology and Biogeography., 11, 377-392.

Ramos-Toro, D. (2018). Historical Conflict and Gender Disparities. Brown University.Richards, P. (1996). Fighting for the rain forest: War, youth and resources in Sierra

Leone. The International African Institute.Ross, M. (2006). A closer look at oil, diamonds and civil war. Annual Review of Political

Science, 9, 265-300.Ross, M. (2008). Oil, Islam and women. American Political Science Review , 102, 107-123.

http://research.jisao.washington.edu/data/ud/

http://research.jisao.washington.edu/data/ud/

https://www.paho.org/mex/index.php?option=com_content&view=article&id=204:antologia-atencion-salud-mexico&Itemid=315

https://www.paho.org/mex/index.php?option=com_content&view=article&id=204:antologia-atencion-salud-mexico&Itemid=315


Rustad, S. A., & Binningsbø, H. M. (2012). A price worth fighting for? Natural resourcesand conflict recurrence. Journal of Peace Research, 49, 531-546.

Sachs, J. D., & Warner, A. M. (1999). The Big Rush, Natural Resource Booms andGrowth. Journal of Development Economics, 59, 43-76.

Sarkees, M. and Wayman, F. (2010). Resort to War: 1816 - 2007 Washington DC: CQPress.

Scheffran, J., Brzoska, M., Kominek, J., Link, P., & Schilling, J. (2012). Climate changeand violent conflict. Science, 336, 869-871.

Tir, J. and Diehl, P. F. (1998). Demographic pressure and interstate conflict: Linkingpopulation growth and density to militarized disputes and wars, 1930-89. Journal ofPeace Research, 35, 319-339.

Sellars, E. & Alix-Garcia, J. (2018). Locational fundamentals, trade, and thechanging urban landscape of Mexico Annual Meeting Agricultural and AppliedEconomics Association, Washington, D.C. https://EconPapers.repec.org/RePEc:ags:aaea18:274238

United Nations (1949). 1948 Demographic Yearbook. Lake Success, NY: United Nations.United Nations (2015). World Population Prospects: The 2015 Revision. Volume 1,

Comprehensive Tables. New York: United Nations. Available at https://esa.un.org/unpd/wpp/Publications/Files/WPP2015_Volume-I_Comprehensive-Tables.pdf,last accessed March 24 2017.

Urdal, H. (2005). People vs. Malthus: population pressures, environmental degradationand armed conflict revisited. Journal of Peace Research, 42, 417-434.

Urdal, H. (2006). A Clash of Generations? Youth Bulges and Political Violence.International Studies Quarterly , 50, 607-629.

Urlanis, B. (2003). Wars and population. Honolulu, Hawaii: University Press of thePacific.

Wooldridge, J. (2002). Econometric analysis of cross section and panel data. Cambridge,MA: MIT Press.

https://EconPapers.repec.org/RePEc:ags:aaea18:274238

https://EconPapers.repec.org/RePEc:ags:aaea18:274238

https://esa.un.org/unpd/wpp/Publications/Files/WPP2015_Volume-I_Comprehensive-Tables.pdf

https://esa.un.org/unpd/wpp/Publications/Files/WPP2015_Volume-I_Comprehensive-Tables.pdf


Figure 1

Reduced Forms, Falsification and First Stages

MEX

ARGAUS

AUT

BGD

BEL

BOL

BRA

BGR

CANCHL

CHN

COL

CRI

CZE

DNK

ECU

SLV

FINFRA

DEU

GRC

GTM HND

HUN

IND IDN

IRLITA

KORMYS

MMR

NLD

NZL

NIC

NOR

PAKPAN PRYPER

PHL

POLPRT

ROMRUSESP

LKA

SWE

CHE

THA

TUR

GBR

USA

URY

VEN

0.5

11

.5

−1.2 −1 −.8 −.6 −.4 −.2Change in predicted mortality, 1940−1980

Base Sample

1a. Change in log population 1940−1980

MEX

ARG

AUT

BGDBOL

BRA

BGR

CHL

CHN

COL

CRI

CZE

ECU

SLV

FINFRAGRC

GTM HND

HUN

IND IDN

IRLITA

KORMYS

MMR

NIC

NOR

PAKPAN PRYPER

PHL

POLPRT

ROMRUSESP

LKA

THA

TUR

URY

VEN

0.5

11

.5


Low and Middle Income Countries

1b. Change in log population 1940−1980

MEXARGAUSBELBOL BRA BGR CANCHL

CHN

COLCRICZE DNKECU

SLV

FINFRA DEU

GRC

GTM

HND HUNIND IDN

IRLITA KOR

MMR

NLDNZL

NIC

NORPAK PANPRY

PER

PHL

POL

PRTROM

RUS

ESP

LKA

SWECHETHA

TUR

GBR USAURYVEN

−1

−.5

0.5

1


Base Sample

1c. Change in fraction of decade in conflict, 1940−1980

MEXARGBOL BRA BGRCHL

CHN

COLCRICZEECU

SLV

FINFRA

GRC

GTM

HND HUNIND IDN

IRLITA KOR

MMR

NIC

NORPAK PANPRY

PER

PHL

POL

PRTROM

RUS

ESP

LKA

THA

TUR

URYVEN

−1

−.5

0.5

1



1d. Change in fraction of decade in conflict, 1940−1980

MEXARGBELBOL BRA BGRCHL

CHN

COL

DNKECUSLV FRA DEU

GRC

GTM HNDITA NLDNIC NORPANPRY

PER PRTROM

RUS

ESP SWECHETHA

TUR

GBR USAURY

VEN

−.5

0.5

1


Base Sample

1e. Change in fraction of decade in conflict 1900−1940

MEXARGBOL BRA BGRCHL

CHN

COL

ECUSLV FRA

GRC

GTM HNDITA NIC NORPANPRY

PER PRTROM

RUS

ESPTHA

TUR

URY

VEN

−.5

0.5

1



1f. Change in fraction of decade in conflict 1900−1940

Notes: See Appendix Table A-1 for variable definition and sources. Initially rich, initially middle

and initially poor are the top, middle and bottom third set of countries in the base sample by level

of income per capita.


Figure 2

Mexico: Reduced Form and First Stage

.2.3

.4.5

.6.7

−1 −.8 −.6 −.4 −.2 0Change in predicted mortality (Mexico), 1940−1960

Base Sample, Binned

2a. Change in log Population, 1940−1960

02

46

8

−1 −.8 −.6 −.4 −.2 0Change in predicted mortality (Mexico), 1940−1960

Base Sample, Binned

2b. Change in Violent Protests, 1940−1960

Notes: See Appendix Table A-1 for variable definition and sources. Points represent averages

within equal size bins constructed from the malaria suitability distribution. The fitted line

shows the regression estimate.


TABLE 1

Descriptive Statistics

By initial income

By declinesin predicted mortality

Year Base sample Rich Middle Poor Abovemedian

Belowmedian

Variable (1) (2) (3) (4) (5) (6) (7)

Fraction of decade in conflict, COW 1900 0.029 0 0.050 0.011 0.047 0.006(0.099) (0) (0.134) (0.033) (0.131) (0.025)



Fraction of decade in conflict, UCDP/PRIO 1940 0.126 0 0.115 0.229 0.201 0.048(0.306) (0) (0.276) (0.417) (0.369) (0.200)

Fraction of decade in conflict, UCDP/PRIO 1980 0.269 0.091 0.190 0.440 0.400 0.134(0.408) (0.302) (0.341) (0.468) (0.447) (0.319)

Fraction of decade in conflict, Fearon-Laitin 1940 0.121 0 0.100 0.237 0.163 0.077(0.302) (0) (0.255) (0.427) (0.347) (0.247)

Fraction of decade in conflict, Fearon-Laitin 1980 0.242 0.091 0.183 0.376 0.367 0.113(0.413) (0.302) (0.369) (0.475) (0.467) (0.306)

log (1 + battle deaths/population) 1940 0.157 0 0.251 0.112 0.208 0.105(0.527) (0) (0.707) (0.305) (0.644) (0.373)

log (1 + battle deaths/population) 1980 0.282 0.002 0.159 0.638 0.541 0.005(0.779) (0.005) (0.528) (1.153) (1.024) (0.017)

Fraction of decade in conflict, natural-resource 1940 0.019 0 0 0.062 0.037 0(0.137) (0) (0) (0.250) (0.192) (0)

Fraction of decade in conflict, natural-resource 1980 0.197 0 0.100 0.396 0.327 0.062(0.397) (0) (0.300) (0.495) (0.471) (0.246)

Fraction of decade in conflict, non-resource 1940 0.135 0 0.115 0.260 0.182 0.087(0.309) (0) (0.276) (0.417) (0.339) (0.273)

Fraction of decade in conflict, non-resource 1980 0.125 0.091 0.103 0.164 0.155 0.094(0.281) (0.302) (0.265) (0.296) (0.305) (0.254)

log population 1900 8.610 8.922 8.238 8.952 8.299 8.909(1.617) (1.410) (1.394) (2.016) (1.668) (1.541)

log population 1940 9.136 9.349 8.738 9.557 9.010 9.261(1.455) (1.344) (1.192) (1.756) (1.530) (1.393)

log population 1980 9.812 9.762 9.393 10.321 9.856 9.768(1.384) (1.293) (1.238) (1.461) (1.484) (1.294)

Baseline predicted mortality 1940 0.469 0.171 0.487 0.626 0.690 0.241(0.271) (0.050) (0.224) (0.272) (0.195) (0.080)

Notes: The table reports the mean values of variables in the samples described in the column heading,with standard deviations in parentheses. Initially rich countries had log GDP per capita over 8.4 in1940, middle-income countries had log GDP per capita between 7.37 and 8.4 and low-income countrieshad log GDP per capita below 7.37 in 1940. Predicted mortality is measured per 100 per year. Columns6 and 7 report descriptive statistics for subsamples in which change in predicted mortality between 1940and 1980s was above or below the median value in the base sample (-0.405). Initially rich countries haveno civil wars recorded in the COW dataset in the 1940s and 1980s, and no conflict incidence accordingto the Fearon and Laitin and UCDP/PRIO sources in the 1940s. See the text and Appendix Table A-1for details and definitions.


TABLE 2

Population and Conflict: OLS Estimates

Dependent variable... Fraction of decade in conflict

COW UCDP/PRIOFearon

& Laitin

log(1+battledeaths/pop. 1940)

(1) (2) (3) (4)

Panel A: Long differences, just 1940s and 1980s

log population 0.323 0.271 0.236 0.722(0.116) (0.135) (0.139) (0.401)

Observations 102 104 104 104R-squared 0.177 0.145 0.098 0.102Number of clusters 50 51 51 51

Panel B: Panel regressions, 1940s-1980s

log population 0.268 0.311 0.251 0.738(0.095) (0.132) (0.132) (0.375)


Panel C: Long differences controlling for age structure, just 1940s and 1980s

log population 0.391 0.316 0.344 1.043(0.140) (0.149) (0.162) (0.469)

Share of population 15-34 -0.995 -0.529 -3.504 -4.852(1.271) (1.544) (2.805) (4.678)


Panel D: Panel regressions controlling for age structure, 1940s-1980s

log population 0.331 0.313 0.265 0.968(0.124) (0.153) (0.148) (0.442)

Share of population 15-34 -1.068 -1.050 -2.095 -4.102(0.490) (0.599) (1.432) (1.858)


Notes: OLS regressions with a full set of year and country fixed effects (equation(4.4) in the text). Robust standard errors (clustered by country) are reportedin parentheses. Panels A and C are long-difference specifications with twoobservations per country, one for the initial date and one for the final date. PanelsB and D are unbalanced panels with one observation per decade. In computingstandard errors, Bangladesh, India and Pakistan are considered a single cluster.See the text and Appendix Table A-1 for definitions and details.

ACEMOGLU ET AL. POPULATION AND CONFLICT 35T

AB

LE

3

Fir

stS

tage

s,R

edu

ced

Form

san

dF

als

ifica

tion

Exe

rcis

es

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Bas

eSam

ple

Low

and

Mid

dle

Inco

me

Countr

ies

Only

Bas

eSam

ple

Low

and

Mid

dle

Inco

me

Countr

ies

Only

Base

Sam

ple

Low

and

Mid

dle

Inco

me

Cou

ntr

ies

Only

Base

Sam

ple

Low

and

Mid

dle

Inco

me

Countr

ies

Only

Fir

stSta

ges

Red

uce

dF

orm

sF

alsi

fica

tion

Exer

cise

s

Dep

enden

tV

aria

ble

is:

log

pop

ula

tion

chan

ge19

40s-

1980s

Fra

ctio

nof

dec

ade

inco

nflic

tch

ange

1940s-

1980s

Fra

ctio

nof

dec

ade

inco

nflic

tch

ange

1900s

-1940s

log

popula

tion

change

190

0s-

1940

s

Chan

gein

pre

dic

ted

mor

tality

,19

40-1

980

-0.7

82-0

.764

-0.4

91

-0.6

60

0.085

0.1

97

-0.1

89

-0.1

98

(0.1

41)

(0.1

91)

(0.1

79)

(0.2

36)

(0.0

55)

(0.1

26)

(0.1

38)

(0.1

96)

Obse

rvat

ions

5140

52

4136

28

52

41

R-s

quar

ed0.

306

0.238

0.1

66

0.2

00

0.0

120.0

39

0.0

33

0.029

Note

s:O

LS

regre

ssio

ns

(equ

ati

on

(4.7

)in

the

text)

.R

ob

ust

stan

dard

erro

rs(c

lust

ered

by

cou

ntr

y)

are

rep

ort

edin

pare

nth

eses

.In

com

pu

tin

gst

an

dard

erro

rs,

Ban

gla

des

h,

Ind

iaan

dP

akis

tan

are

con

sid

ered

asi

ngle

clu

ster

.S

eeth

ete

xt

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.


TABLE 4

The Effect of Population on Conflict: 2SLS Estimates

Dependent variable is fraction of decade in conflict(cols 1-3), and log (1 + battle deaths/pop. in 1940)

in col 4

(1) (2) (3) (4)

COW UCDP/PRIO Fearon &log(1+battle

deaths/Laitin pop. 1940)

Panel A: Long differences, just 1940s and 1980s

log population 0.617 0.576 0.879 1.347(0.213) (0.238) (0.303) (0.598)

Observations 102 104 104 104Number of clusters 50 51 51 51

Panel B: Panel regressions, 1940s-1980s

log population 0.609 0.304 0.873 1.106(0.205) (0.250) (0.461) (0.454)

Observations 307 308 308 273Number of clusters 63 63 63 54

Notes: 2SLS regressions with a full set of year and country fixed effects (equation (4.4)in the text, where population is instrumented by predicted mortality, as in equation(4.6) in the text). Robust standard errors (clustered by country) are reported inparentheses. Panel A presents long-difference specifications with two observationsper country, one for the initial date and one for the final date. Panel B presentsunbalanced panels with one observation per decade. First stage, for the sample withdata on years in conflict according to COW, in column 1 (panel B and D) of Table5. In computing standard errors, Bangladesh, India and Pakistan are considered asingle cluster. See the text and Appendix Table A-1 for definitions and details.


AB

LE

5

The

Eff

ect

of

Popu

lati

on

on

Con

flic

t:2S

LS

an

dF

irst

-Sta

geE

stim

ate

sD

epen

den

tva

riable

isfr

act

ion

of

dec

ade

inco

nflic

tacc

ordin

gto

Cor

rela

tes

ofW

ar

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Bas

eSam

ple

Excl

udin

gE

aste

rnE

uro

pe

Excl

udin

gW

este

rnE

uro

pe

Low

and

Mid

dle

Inco

me

Cou

ntr

ies

Only

Excl

udin

gM

ost

Aff

ecte

dB

yW

WII

Base

Sam

ple

Ass

ign

1950

to1940

Base

Sam

ple

Ass

ign

194

6-4

9to

1940

Addin

gP

opula

tion

15-

34

asC

ovar

iate

Addin

gL

ife

Exp

ecta

ncy

at20

asC

ovar

iate

Just

1940

san

d20

00s

Pan

elA

:2SL

Ses

tim

ate

s.D

epen

den

tva

riable

isfr

acti

onof

dec

ade

inco

nflic

tac

cord

ing

toC

orr

elate

sof

War

(Long

diff

eren

ces,

just

1940s

and

198

0s)

log

pop

ula

tion

0.6

17

0.6

57

0.9

85

0.8

280.

610

0.420

0.6

40

0.7

040.

987

0.2

96(0

.213

)(0

.240

)(0

.363)

(0.2

94)

(0.2

15)

(0.1

98)

(0.2

26)

(0.2

28)

(0.5

50)

(0.1

44)

Pan

elB

:F

irst

-sta

gees

tim

ates

.D

epen

den

tva

riab

leis

log

pop

ula

tion

(Long

diff

eren

ces,

just

1940

sand

1980

s)

Bas

elin

epre

dic

ted

mort

alit

y-0

.782

-0.7

00-0

.584

-0.7

64

-0.8

11

-0.7

82-0

.782

-0.8

48

-0.5

44

-1.1

74(0

.141

)(0

.141

)(0

.168)

(0.1

91)

(0.1

40)

(0.1

41)

(0.1

41)

(0.1

73)

(0.2

33)

(0.1

90)

R-s

quar

ed0.

823

0.84

70.8

87

0.8

28

0.8

42

0.8

23

0.823

0.83

20.

854

0.815

Obse

rvat

ions

(Pan

els

Aan

dB

)10

292

72

8094

102

102

8684

102

Num

ber

ofcl

ust

ers

(Pan

els

Aan

dB

)50

4535

39

46

5050

4341

50

Pan

elC

:2S

LS

esti

mat

es.

Dep

enden

tva

riab

leis

frac

tion

of

dec

ade

inco

nflic

tacc

ord

ing

toC

orr

elate

sof

War

(Panel

regr

essi

ons,

1940s

-1980

s)

log

pop

ula

tion

0.6

09

0.6

49

1.0

36

0.8

620.

615

0.276

0.6

49

0.5

261.

020

0.4

67(0

.205

)(0

.241

)(0

.406)

(0.3

07)

(0.2

11)

(0.1

10)

(0.2

36)

(0.2

12)

(0.4

64)

(0.1

52)

Pan

elD

:F

irst

-sta

gees

tim

ates

.D

epen

den

tva

riab

leis

log

pop

ula

tion

(Pan

elre

gre

ssio

ns,

1940

s-1980s

)

Bas

elin

epre

dic

ted

mort

alit

y-0

.464

-0.4

02-0

.336

-0.4

71

-0.4

76

-0.4

64-0

.464

-0.4

82

-0.3

15

-0.6

79(0

.094

)(0

.093

)(0

.118)

(0.1

31)

(0.0

95)

(0.0

94)

(0.0

94)

(0.1

18)

(0.1

40)

(0.1

21)

R-s

quar

ed0.

792

0.81

90.8

85

0.8

14

0.8

15

0.7

92

0.792

0.75

70.

810

0.774

Obse

rvat

ions

(Pan

els

Can

dD

)30

727

822

8252

279

307

307

227

202

437

Num

ber

ofcl

ust

ers

(Pan

els

Cand

D)

63

5747

5257

6363

4648

63

Note

s:P

an

els

Aan

dC

:2S

LS

regre

ssio

ns

wit

ha

full

set

of

yea

ran

dco

untr

yfi

xed

effec

ts(e

qu

ati

on

(4.4

)in

the

text,

wh

ere

pop

ula

tion

isin

stru

men

ted

by

pre

dic

ted

mort

ality

,as

ineq

uati

on

(4.6

)in

the

text)

.P

an

els

Ban

dD

:O

LS

regre

ssio

ns

wit

ha

full

set

of

yea

ran

dco

untr

yfi

xed

effec

ts(e

qu

ati

on

(4.6

)in

the

text)

.R

ob

ust

stan

dard

erro

rs(c

lust

ered

by

cou

ntr

y)

are

rep

ort

edin

pare

nth

eses

.P

an

els

Aan

dC

pre

sent

lon

g-d

iffer

ence

spec

ifica

tion

sw

ith

two

ob

serv

ati

on

sp

erco

untr

y,on

efo

rth

ein

itia

ld

ate

an

don

efo

rth

efi

nal

date

.P

an

els

Ban

dD

pre

sent

unb

ala

nce

dp

an

els

wit

hon

eob

serv

ati

on

per

dec

ad

e.In

com

pu

tin

gst

an

dard

erro

rs,

Ban

gla

des

h,

Ind

iaan

dP

akis

tan

are

con

sid

ered

asi

ngle

clu

ster

.S

eeth

ete

xt

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.

38 REVIEW OF ECONOMIC STUDIEST

AB

LE

6

The

Eff

ect

of

Popu

lati

on

on

Con

flic

t:2S

LS

an

dF

irst

-Sta

geE

stim

ate

sR

obu

stn

ess

toD

iffer

enti

al

Tre

nds

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Inte

ract

ion

of

post

-yea

rd

um

mie

sw

ith

...

Inst

itu

tion

slo

gG

DP

per

cap

ita

in193

0

log

Pop

ula

tion

in19

30

Inte

ract

ion

ofS

hare

Pop

ula

tion

15-

34an

d20-

39,

194

0

Dia

mon

dP

rod

uct

ion

per

Cap

ita

in196

0

Oil

and

Gas

Ren

tsp

erC

ap

ita

in19

60

Eth

nic

Pol

ariz

atio

nIn

itia

lW

arin

1930

sL

atit

ud

eM

alar

iaE

colo

gyIn

dex

Pan

elA

:2S

LS

esti

mate

s.D

epen

den

tva

riab

leis

frac

tion

ofdec

ade

inco

nfl

ict

acc

ord

ing

toC

orre

late

sof

War

(Lon

gd

iffer

ence

s,ju

st19

40s

an

d198

0s)

log

pop

ula

tion

0.74

81.1

32

0.6

170.7

220.6

23

0.596

0.738

0.49

70.

864

0.73

2(0

.257

)(0

.427

)(0

.215

)(0

.236

)(0

.213)

(0.2

04)

(0.2

94)

(0.2

24)

(0.3

44)

(0.2

58)

p-v

alu

efo

rp

ost

year

du

mm

yx

vari

ab

lein

dic

ated

atth

eto

pof

each

colu

mn

0.106

0.0

541

0.973

0.672

<0.

010.0

173

0.4

010.

533

0.11

10.

0106

Pan

elB

:F

irst

-sta

gees

tim

ates

.D

epen

den

tva

riab

leis

log

pop

ula

tion

(Lon

gd

iffer

ence

s,ju

st19

40s

and

1980

s)

Base

lin

ep

red

icte

dm

orta

lity

-0.7

30

-0.6

00-0

.762

-0.8

36

-0.7

76-0

.808

-0.5

99-0

.935

-0.5

40

-0.7

04(0

.167

)(0

.219

)(0

.126

)(0

.183

)(0

.138)

(0.1

38)

(0.1

48)

(0.1

89)

(0.1

43)

(0.1

42)

R-s

qu

are

d0.

826

0.8

33

0.844

0.800

0.84

20.8

41

0.868

0.80

50.

860

0.83

3p

-valu

efo

rp

ost

year

du

mm

yx

vari

ab

lein

dic

ated

atth

eto

pof

each

colu

mn

0.461

0.0

952

<0.

010.

823

<0.

01<

0.01

<0.

01

0.93

1<

0.01

0.07

01

Ob

serv

atio

ns

(Pan

els

Aan

dB

)10

2100

102

86

102

102

9688

102

102

Nu

mb

erof

clu

ster

s(P

an

els

Aan

dB

)50

4950

4350

5047

4450

50

Pan

elC

:2S

LS

esti

mate

s.D

epen

den

tva

riab

leis

frac

tion

ofd

ecad

ein

con

flic

tacc

ord

ing

toC

orre

late

sof

War

(Pan

elre

gres

sion

s,194

0s-

1980

s)

log

pop

ula

tion

0.70

61.1

50

0.6

170.5

170.6

14

0.592

0.712

0.32

80.

835

0.70

9(0

.252

)(0

.457

)(0

.211

)(0

.206

)(0

.204)

(0.1

97)

(0.3

03)

(0.1

35)

(0.3

49)

(0.2

33)

p-v

alu

efo

rp

ost

year

du

mm

yx

vari

ab

lein

dic

ated

atth

eto

pof

each

colu

mn

0.165

0.0

640

0.016

00.

647

<0.

010.0

958

0.5

620.

0480

0.23

60.

101

Pan

elD

:F

irst

-sta

gees

tim

ate

s.D

epen

den

tva

riab

leis

log

pop

ula

tion

(Pan

elre

gre

ssio

ns,

194

0s-1

980s

)

Base

lin

ep

red

icte

dm

orta

lity

-0.4

37

-0.3

61-0

.446

-0.4

95

-0.4

60-0

.478

-0.3

37-0

.559

-0.3

06

-0.4

11(0

.113

)(0

.146

)(0

.086

)(0

.120

)(0

.093)

(0.0

92)

(0.0

96)

(0.1

26)

(0.0

92)

(0.0

94)

R-s

qu

are

d0.

810

0.8

18

0.798

0.750

0.80

10.8

09

0.849

0.75

70.

844

0.80

6p

-valu

efo

rp

ost

year

du

mm

yx

vari

ab

lein

dic

ated

atth

eto

pof

each

colu

mn

<0.

01<

0.01

0.0

538

0.0

701

<0.

01<

0.01

<0.

01

0.05

29<

0.01

0.03

76

Ob

serv

atio

ns

(Pan

els

Can

dD

)30

7267

265

223

300

300

277

244

300

307

Nu

mb

erof

clu

ster

s(P

an

els

Can

dD

)63

5352

4561

6156

5061

63

Note

s:P

an

els

Aan

dC

:2S

LS

regre

ssio

ns

wit

ha

full

set

of

yea

ran

dco

untr

yfi

xed

effec

ts(e

qu

ati

on

(4.9

)in

the

text,

wh

ere

pop

ula

tion

isin

stru

men

ted

by

pre

dic

ted

mort

ality

,as

ineq

uati

on

(4.6

)in

the

text)

.P

an

els

Ban

dD

:O

LS

regre

ssio

ns

wit

ha

full

set

of

yea

ran

dco

untr

yfi

xed

effec

ts(e

qu

ati

on

(4.8

)in

the

text)

.R

ob

ust

stan

dard

erro

rs(c

lust

ered

by

cou

ntr

y)

are

rep

ort

edin

pare

nth

eses

.P

an

els

Aan

dB

pre

sent

lon

g-d

iffer

ence

spec

ifica

tion

sw

ith

two

ob

serv

ati

on

sp

erco

untr

y,on

efo

rth

ein

itia

ld

ate

an

don

efo

rth

efi

nal

date

.P

an

els

Can

dD

pre

sent

unb

ala

nce

dp

an

els

wit

hon

eob

serv

ati

on

per

dec

ad

e.In

com

pu

tin

gst

an

dard

erro

rs,

Ban

gla

des

h,

Ind

iaan

dP

akis

tan

are

con

sid

ered

asi

ngle

clu

ster

.S

eeth

ete

xt

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.


AB

LE

7

The

Eff

ect

of

Popu

lati

on

on

Con

flic

t:C

on

troll

ing

Fle

xibl

yfo

rth

eIm

pact

of

Init

ial

Con

flic

tan

dM

ean

Rev

ersi

on

,U

sin

gP

an

elD

ata

Dep

end

ent

vari

ab

leis

fract

ion

of

dec

ad

ein

con

flic

tacc

ord

ing

toC

orr

elate

sof

War

Bas

elin

eS

pec

ifica

tion

Incl

ud

ing

Init

ial

War

in194

0,

Inte

ract

edw

ith

Tim

eD

um

mie

sIn

clu

din

gIn

itia

lP

op

.in

1940,

Inte

ract

edw

ith

Tim

eD

um

mie

sN

on

lin

ear

Est

imate

s2S

LS

2S

LS

2SL

S2S

LS

GM

M2S

LS

2S

LS

GM

MG

MM

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

Init

ial

War

isfr

acti

on

ofd

ecad

ein

con

flic

tin

the

1940

s

log

pop

ula

tion

0.60

90.6

060.5

84

0.2

66

0.2

38

0.6

09

0.3

04

0.1

58

(0.2

05)

(0.2

07)

(0.1

81)

(0.1

07)

(0.1

05)

(0.2

07)

(0.1

43)

(0.0

84)

Lag

ged

con

flic

t-0

.002

0.1

57

-0.0

49

0.1

80

(0.0

82)

(0.1

14)

(0.0

73)

(0.1

54)

π0.4

45

(0.0

38)

λ-0

.062

(0.0

08)

Ob

serv

atio

ns

307

281

281

272

235

265

254

245

Nu

mb

erof

clu

ster

s63

56

56

56

52

52

Eff

ect

ofan

incr

ease

of0.

10in

log

pop

from

194

0to

1980

0.06

10.0

61

0.0

58

0.0

27

0.0

28

0.0

61

0.0

29

0.0

19

0.0

45

p-v

alu

efo

rp

ost

year

du

mm

yx

init

ial

war

<0.0

1<

0.0

10.3

03

0.0

13

0.06

<0.0

1M

omen

ts38

38

89

Han

sen

p-v

alu

e0.3

92

0.7

21

0.9

88

AR

2p

-valu

e0.5

25

0.9

79

Note

s:U

nb

ala

nce

dp

an

els

wit

hon

eob

serv

ati

on

per

dec

ad

e.C

olu

mn

1is

ou

rb

ase

lin

e2S

LS

regre

ssio

ns

wit

ha

full

set

of

yea

ran

dco

untr

yfi

xed

effec

ts(e

qu

ati

on

(4.4

)in

the

text,

wh

ere

pop

ula

tion

isin

stru

men

ted

by

pre

dic

ted

mort

ality

,as

ineq

uati

on

(4.6

)in

the

text)

.C

olu

mn

2re

stri

cts

the

sam

ple

toth

ese

tof

cou

ntr

ies

for

wh

ich

ther

eis

availab

led

ata

on

“in

itia

l”(i

nth

e1940s)

con

flic

t,m

easu

red

as

the

fract

ion

of

dec

ad

ein

con

flic

t.C

olu

mn

s3,

4an

d5

incl

ud

ea

full

set

of

yea

rd

um

mie

sin

tera

cted

wit

hin

itia

lw

ar.

Colu

mn

s6,

7an

d8

inte

ract

inst

ead

wit

hin

itia

lp

op

ula

tion

.T

he

Are

llan

oan

dB

on

d’s

GM

Mes

tim

ato

rs(c

olu

mn

s5

an

d8)

rem

ove

cou

ntr

yfi

xed

effec

tsby

takin

gfi

rst

diff

eren

ces

an

dth

enco

nst

ruct

mom

ent

con

dit

ion

su

sin

gall

pre

det

erm

ined

lags

of

con

flic

tan

dp

red

icte

dm

ort

ality

as

inst

rum

ents

.It

ises

tim

ate

din

two

step

san

dis

op

tim

all

yw

eighte

d.

Colu

mn

9p

rese

nts

the

wei

ghte

dtw

o-s

tep

GM

Mes

tim

ate

sof

the

Blo

om

etal.

(2014)

pro

pose

deq

uati

on

(3):

∆c it

=αt

+π

∆lo

gp

op

ula

tionit

+λπ

log

pop

ula

tioni,t−

1−λc i,t−1

+ε it.

Inco

lum

n9,

the

seco

nd

an

dall

lon

ger

lags

of

con

flic

tan

dth

efi

rst

an

dall

lon

ger

lags

of

pre

dic

ted

mort

ality

are

use

das

inst

rum

ents

.R

ob

ust

stan

dard

erro

rsco

rrec

ted

for

arb

itra

ryse

rial

corr

elati

on

clu

ster

edat

the

cou

ntr

yle

vel

(Ban

gla

des

h,

Ind

iaan

dP

akis

tan

are

con

sid

ered

asi

ngle

clu

ster

)are

rep

ort

edin

colu

mn

s1-4

an

d6-7

,an

dro

bu

stst

an

dard

erro

rsare

rep

ort

edin

colu

mn

s5,

8an

d9.

See

the

text

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.


AB

LE

8

The

Eff

ect

of

Popu

lati

on

on

Con

flic

t:T

he

Impo

rtan

ceof

Natu

ral-

Res

ou

rce

Rel

ate

dC

on

flic

ts,

Popu

lati

on

Den

sity

,E

con

om

icG

row

than

dE

con

om

icH

ard

ship

:2S

LS

Est

imate

s

Dep

enden

tva

riab

leis

frac

tion

ofdec

ade

in...

GD

Pgr

owth

Neg

ati

vepri

cesh

ock

sN

atura

lR

esou

rce

Con

flic

tN

on-r

esourc

eC

onflic

tH

igh

Pop

.D

ensi

tyL

owP

op.

Den

sity

Con

flic

t slow

Conflic

t fast

Con

flic

t high

hard

ship

Conflic

t low

hard

hsip

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Pan

elA

:L

ong

diff

eren

ces,

just

1940

san

d19

80s

log

ofp

opula

tion

0.91

80.

094

0.93

00.3

300.

620

0.0

16

0.4

53

0.1

64

(0.2

53)

(0.1

57)

(0.3

35)

(0.2

33)

(0.2

11)

(0.0

25)

(0.1

86)

(0.1

67)

Obse

rvat

ions

102

102

5050

5151

102

102

Num

ber

ofcl

ust

ers

5050

2425

5050

50

50

Pan

elB

:P

anel

regr

essi

ons,

1940

s-19

80s

log

ofp

opula

tion

0.69

00.

000

1.01

90.1

620.

456

0.1

42

0.5

81

0.0

26

(0.1

99)

(0.2

26)

(0.3

04)

(0.1

83)

(0.1

58)

(0.0

90)

(0.1

67)

(0.1

37)

Obse

rvat

ions

308

308

134

133

306

306

307

314

Num

ber

ofcl

ust

ers

6363

2627

6363

63

63

Note

s:2S

LS

regre

ssio

ns

wit

ha

full

set

of

yea

ran

dco

untr

yfi

xed

effec

ts(e

qu

ati

on

(4.4

)in

the

text,

wh

ere

pop

ula

tion

isin

stru

men

ted

by

pre

dic

ted

mort

ality

,as

ineq

uati

on

(4.6

)in

the

text)

.R

ob

ust

stan

dard

erro

rsco

rrec

ted

for

arb

itra

ryse

rial

corr

elati

on

clu

ster

edat

the

cou

ntr

yle

vel

(Ban

gla

des

h,

Ind

iaan

dP

akis

tan

are

con

sid

ered

asi

ngle

clu

ster

).N

atu

ral-

reso

urc

ean

dN

on

-res

ou

rce

con

flic

tsas

class

ified

for

the

UC

DP

/P

RIO

con

flic

td

ata

set

by

Sir

i&

Bin

nin

gsb

ø(2

012).

InP

an

elA

,C

on

flic

t slow

iseq

ual

toze

roif

the

chan

ge

inlo

gG

DP

per

cap

ita

exce

eds

the

med

ian

inth

esa

mp

le(f

ast

gro

wer

s)an

deq

ual

toth

ech

an

ge

inco

nfl

ict

from

1940

to1980

(ch

an

ge

inth

efr

act

ion

of

the

dec

ad

ein

con

flic

tacc

ord

ing

toC

orr

elate

sof

War)

ifth

ech

an

ge

inlo

gG

DP

per

cap

ita

issm

aller

than

the

med

ian

inth

esa

mp

le(s

low

gro

wer

s).

Con

flic

t fast

isd

efin

edan

alo

gou

sly

as

equ

al

toth

ech

an

ge

inco

nfl

ict

for

fast

gro

wer

san

dze

rofo

rsl

ow

gro

wer

s.T

hes

eare

regre

ssed

on

chan

ges

inlo

gp

op

ula

tion

from

1940

to1980.

InP

an

elB

,C

on

flic

t slow

iseq

ual

toze

roin

each

dec

ad

eif

the

chan

ge

inlo

gG

DP

per

cap

ita

inth

ed

ecad

eex

ceed

sth

em

edia

nin

the

sam

ple

(fast

gro

wer

s)an

deq

ual

toth

efr

act

ion

of

dec

ad

ein

con

flic

tacc

ord

ing

toC

orr

elate

sof

War

oth

erw

ise

(slo

wgro

wer

s).

Con

flic

t fast

isd

efin

edan

alo

gou

sly

as

equ

al

toco

nfl

ict

for

fast

gro

wer

san

dze

rofo

rsl

ow

gro

wer

s.C

on

flic

t high

hard

ship

an

dC

on

flic

t low

hard

ship

are

con

stru

cted

sim

ilar

toC

on

flic

t slow

an

dC

on

flic

t fast

,b

ut

base

don

wh

eth

erth

enu

mb

erof

yea

rsw

ith

econ

om

ich

ard

ship

isab

ove

or

bel

ow

the

med

ian

of

its

dis

trib

uti

on

du

rin

gth

ere

levant

sam

ple

per

iod

.W

eid

enti

fyti

mes

of

econ

om

ich

ard

ship

as

yea

rsin

wh

ich

an

ind

exof

com

mod

ity

pri

cesh

ock

s(c

om

pu

ted

as

the

sum

of

com

mod

ity

pri

cech

an

ges

wei

ghte

dby

cou

ntr

y-s

pec

ific

exp

ort

top

op

ula

tion

rati

os)

isn

egati

ve.

See

the

text

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.


TABLE 9

Mexico: Descriptive statisticsBy declines in predicted

mortality (Mexico)By declines in predicted

mortality (Mexico)

Year Base sample Abovemedian

Belowmedian

Year Base sample Abovemedian

Belowmedian

Variable (1) (2) (3) (4) Variable (5) (6) (7) (8)

Social Conflict Disease

Violent protests 1960 2.314 2.995 1.637 Baseline predicted mortality (Mexico) 1940 0.426 0.668 0.184(11.801) (13.690) (9.517) (0.282) (0.160) (0.125)

Violent natural resource protests 1960 1.024 1.516 0.535 Droughts(8.849) (11.162) (5.638)

Violent non-resource protests 1960 1.283 1.469 1.099 Non-harvest droughts 1960 7.179 7.114 7.243(7.066) (7.055) (7.074) (5.628) (5.716) (5.541)

Non-violent protests 1960 11.987 12.039 11.935 Harvest droughts 1960 1.994 1.709 2.279(39.049) (34.597) (43.036) (3.136) (2.654) (3.531)

Historical conflict 1860-1940 0.048 0.041 0.055 Education(0.241) (0.206) (0.272)

Population Literacy rate 1940 0.314 0.286 0.343(0.192) (0.190) (0.190)

log population 1920 7.991 7.897 8.083 Literacy rate 1960 0.509 0.469 0.550(1.099) (1.071) (1.119) (0.178) (0.178) (0.168)

log population 1930 8.140 8.054 8.225 Literacy rate, 15-39 1940 0.371 0.334 0.410(1.075) (1.023) (1.118) (0.222) (0.218) (0.219)

log population 1940 8.279 8.188 8.370 Literacy rate, 15-39 1960 0.592 0.542 0.642(1.095) (1.049) (1.132) (0.218) (0.217) (0.207)

log population 1960 8.690 8.645 8.734 Primary school enrollment 1940 0.031 0.026 0.036(1.191) (1.145) (1.234) (0.049) (0.043) (0.055)

Share of population 15-34 1940 0.318 0.324 0.311 Primary school enrollment 1960 0.079 0.066 0.093(0.024) (0.023) (0.023) (0.088) (0.081) (0.093)

Share of population 15-34 1960 0.309 0.312 0.306(0.028) (0.028) (0.027)



Notes: Municipal-level information. The table reports the mean values of variables in the samplesdescribed in the column heading, with standard deviations in parentheses. Predicted mortality (Mexico)is equal to malaria suitability in 1940 and to zero in 1960. Protests are counts of news stories aboutprotests, expressed as a fraction of baseline population (per 100,000 people). Historical conflicts arebattles fought in each municipality from 1816 to 1940. Columns 3 and 4 report descriptive statistics forsubsamples in which malaria suitability was above or below the median value in the base sample (0.39).See the text and Appendix Table A-1 for more details and definitions.


AB

LE

10

Mex

ico:

Fir

stS

tage

,R

edu

ced

Form

an

dF

als

ifica

tion

Exe

rcis

es

Fir

stSta

geR

educe

dF

orm

Fal

sifica

tion

Exer

cise

s

Dep

enden

tva

riab

leis

:lo

gp

opula

tion

change

1940

s-19

60s

Vio

lent

Pro

test

sch

ange

1940s

-196

0slo

gp

opula

tion

chan

ge19

30s-

1940

slo

gp

opula

tion

change

1920

s-19

40s

Sh

are

of

popula

tion

20-3

9ch

ange

1930

s-1940

s

Lit

eracy

rate

,ch

ange

1930

s-194

0sH

isto

rica

lC

onflic

t

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Pan

elA

:N

oco

ntr

ols

∆P

redic

ted

mor

tality

(Mex

ico)

-0.2

87-3

.518

-0.0

43-0

.065

-0.0

01

0.021

-0.0

16

(0.0

42)

(1.2

39)

(0.0

24)

(0.0

46)

(0.0

05)

(0.0

17)

(0.0

20)

Obse

rvat

ions

2,379

2,3

812,

370

2,35

32,0

16

2,0

15

2,4

01

Pan

elB

:C

ontr

ols

∆P

redic

ted

mor

tality

(Mex

ico)

-0.2

13-2

.754

-0.0

010.

023

0.0

05

0.0

13

-0.0

14(0

.044)

(1.1

28)

(0.0

30)

(0.0

47)

(0.0

05)

(0.0

15)

(0.0

29)

Obse

rvat

ions

1,789

1,7

861,

781

1,77

11,7

17

1,7

16

1,7

89

Note

s:M

un

icip

al-

level

OL

Sre

gre

ssio

ns

wit

ha

full

set

of

state

fixed

effec

ts.R

ob

ust

stan

dard

erro

rsallow

ing

for

spati

alco

rrel

ati

on

bet

wee

nm

un

icip

aliti

esw

ith

ina

rad

ius

of

35.9

km

.P

redic

ted

mort

ality

(Mex

ico)

ism

ala

ria

suit

ab

ilit

yin

1940

an

dze

roin

1960.

Pre

dic

ted

mort

ality

(Mex

ico)

iseq

ualto

mala

ria

suit

ab

ilit

yin

1940

an

dto

zero

in1960.

∆d

enote

sch

an

ges

from

1940

to1960.

Contr

ols

are

dis

tan

ceto

cap

ital,

dis

tan

ceto

big

citi

es,

lan

dqu

ality

ind

ex,

log

pop

ula

tion

in1940,

pri

mary

sch

ool

in1940,

un

iver

sity

enro

llm

ent

in1940,

share

of

am

un

icip

ality

’sare

aon

ase

dim

enta

ryb

asi

n,

share

of

ind

igen

ou

s1940

an

dh

isto

rica

lco

nfl

ict

(exce

pt

inco

lum

n9).

See

the

text

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.


TABLE 11

Mexico: Population and Violent Protests per capita 2SLS, First Stage and OLSEstimates

Dependent variable is violent protests (Panels A and C)and log population (Panel B)

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Panel A: 2SLS estimates

log population 12.372 13.333 12.813 13.102 12.389 12.996 14.263 10.948 13.049(3.218) (3.040) (3.069) (3.047) (3.166) (3.056) (3.525) (3.113) (4.130)

Baseline control x post-anti-malaria campaign dummy

log population, 1940 0.621 -0.108(0.220) (0.283)

Primary school, 1940 22.070 22.467(6.101) (7.709)

University enrollment, 1940 161.524 3.572(90.083) (43.801)

Historical conflict 3.021 1.746(0.716) (0.596)

Sedimentary basin -0.617 -0.558(0.567) (0.830)

Share indigenous, 1940 -1.818 -1.029(0.752) (0.913)

Panel B: First-stage estimates

Predicted mortality (Mexico) -0.284 -0.262 -0.259 -0.247 -0.236 -0.259 -0.240 -0.246 -0.211(0.030) (0.030) (0.030) (0.029) (0.030) (0.030) (0.032) (0.028) (0.031)

Panel C: OLS estimates

log population 4.700 4.780 4.700 4.469 5.163 4.577 4.748 6.152 5.132(0.841) (0.853) (0.843) (0.880) (0.914) (0.850) (0.858) (1.033) (1.067)

Observations 4,754 4,754 4,754 4,128 4,176 4,754 4,754 3,760 3,572

Geographic controls xpost-anti-malaria cam-paign dummy

X X X X X X X X

Notes: Municipality-level regressions with observations for 1940 and 1960. All regressions include a fullset of municipality fixed effects as well as state fixed effects interacted with the post year dummy.Additional controls are indicated in Panel A (and also included in the other two panels but not reportedto save space). Robust standard errors allowing for spatial correlation between municipalities within aradius of 35.9 km. Protests are counts of news stories about protests expressed as a fraction of baselinepopulation (per 100,000 people). Predicted mortality (Mexico) is malaria suitability in 1940 and zero in1960. Predicted mortality (Mexico) is equal to malaria suitability in 1940 and to zero in 1960. Geographiccontrols are distance to capital, distance to big cities and land quality. See the text and Appendix TableA-1 for definitions and details.


AB

LE

12

Mex

ico:

Popu

lati

on

an

dP

rote

sts

per

capit

a2S

LS

Est

imate

sH

eter

ogen

eou

sE

ffec

tsan

dO

ther

Ou

tcom

es

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Pan

elA

:2S

LS

esti

mat

esD

epen

den

tva

riab

leis

...

Pro

test

sp

erca

pit

aby

typ

eV

iole

nt

pro

test

sp

erca

pit

a

Vio

lent

Nat

ura

lN

atu

ral

Non-

Non

-N

onvio

lent

Non

vio

lent

Inte

ract

ions

wit

hR

esou

rce

Res

ourc

eR

esourc

eR

esou

rce

Non

-harv

est

&H

arv

est

Dro

ughts

log

pop

ula

tion

7.62

18.

737

4.7

514.3

13

27.4

61

22.3

4310.

771

14.9

98

10.4

46

16.6

81(2

.667

)(2

.839)

(1.6

30)

(2.6

86)

(8.7

85)

(13.9

28)

(3.6

23)

(5.0

45)

(3.9

24)

(4.9

87)

log

pop×

Non

-har

vest

0.4

46

0.5

42

1.173

1.681

(0.1

50)

(0.1

75)

(0.6

79)

(0.7

51)

log

pop×

Har

vest

-0.2

95-0

.165

-1.7

85

-2.1

17

(0.3

45)

(0.4

58)

(1.3

18)

(1.5

19)

Pan

elB

:F

irst

-sta

gees

tim

ates

Dep

enden

tva

riab

leis

log

pop

ula

tion

Pre

dic

ted

mor

tality

(Mex

ico)

-0.2

84

-0.2

11

-0.2

84

-0.2

11

-0.2

84

-0.2

11

-0.2

45

-0.2

04

-0.1

73

-0.1

57

(0.0

30)

(0.0

31)

(0.0

30)

(0.0

31)

(0.0

30)

(0.0

31)

(0.0

37)

(0.0

34)

(0.0

46)

(0.0

50)

Pre

dic

ted

mor

tality

(Mex

ico)×

Non

-har

vest

-0.0

07

-0.0

02

-0.0

16

-0.0

07

(0.0

03)

(0.0

03)

(0.0

05)

(0.0

05)

Pre

dic

ted

mor

tality

(Mex

ico)×

Har

vest

0.0

14

0.0

070.0

16

0.0

07

(0.0

06)

(0.0

07)

(0.0

10)

(0.0

11)

Obse

rvat

ions

4,75

93,5

72

4,7

59

3,572

4,7

56

3,5

724,

744

3,5

66

4,7

443,5

66

Con

trol

sx

pos

t-an

ti-m

alar

iaca

mpai

gndum

my

XX

XX

XD

rough

tsx

pos

t-an

ti-m

alar

iaca

mpai

gndum

my

XX

Note

s:M

un

icip

ality

-lev

elre

gre

ssio

ns

wit

hob

serv

ati

on

sfo

r1940

an

d1960.

All

regre

ssio

ns

incl

ud

ea

full

set

of

mu

nic

ipality

fixed

effec

tsas

wel

las

state

fixed

effec

tsin

tera

cted

wit

hth

ep

ost

yea

rd

um

my.

Pop

ula

tion

isin

stru

men

ted

wit

hp

red

icte

dm

ort

ality

(Mex

ico).

Rob

ust

stan

dard

erro

rsallow

ing

for

spati

al

corr

elati

on

bet

wee

nm

un

icip

aliti

esw

ith

ina

rad

ius

of

35.9

km

.P

rote

sts

are

cou

nts

of

new

sst

ori

esab

ou

tp

rote

sts

in1960,

exp

ress

edas

afr

act

ion

of

base

lin

ep

op

ula

tion

(per

100,0

00

peo

ple

).P

red

icte

dm

ort

ality

(Mex

ico)

ism

ala

ria

suit

ab

ilit

yin

1940

an

dze

roin

1960.

Pre

dic

ted

mort

ality

(Mex

ico)

iseq

ual

tom

ala

ria

suit

ab

ilit

yin

1940

an

dto

zero

in1960.

Ad

rou

ght

isd

efin

edas

pre

cip

itati

on

bel

ow

the

5th

per

centi

leof

the

lon

g-r

un

dis

trib

uti

on

(1900-2

008)

of

month

lyra

inp

erm

un

icip

ality

an

dw

eco

unt

the

nu

mb

erof

month

sw

ith

dro

ughts

inth

ed

ecad

e.C

ontr

ols

are

dis

tan

ceto

cap

ital,

dis

tan

ceto

big

citi

es,

lan

dqu

ality

ind

ex,

log

pop

ula

tion

in1940,

pri

mary

sch

ool

in1940,

un

iver

sity

enro

llm

ent

in1940,

share

of

am

un

icip

ality

’sare

aon

ase

dim

enta

ryb

asi

n,

share

of

ind

igen

ou

s1940

an

dh

isto

rica

lco

nfl

ict.

Dro

ughts

are

dem

ean

edb

efore

inte

ract

ion

.S

eeth

ete

xt

an

dA

pp

end

ixT

ab

leA

-1fo

rd

efin

itio

ns

an

dd

etails.

population and con ict - the review of economic studies€¦ · ict might work through channels...

Documents